ISSN: 1402-1757 ISBN 978-91-7439-XXX-X Se i listan och fyll i siffror där kryssen är

LICENTIATE T H E SIS

Pooya Vahdati Identification of Soil Parameters in an Embankment Dam by Mathematical Optimization

Department of Civil, Environmental and Natural Resources Engineering

Division of Mining and Geotechnical Engineering LICENTIATE THESIS

ISSN 1402-1757 Identification of Soil Parameters ISBN 978-91-7439-852-6 (print) ISBN 978-91-7439-853-3 (pdf) inIdentification an Embankment of soil parameters Dam in byan embankment dam by mathematical optimization Luleå University of Technology 2014 Mathematical Optimization

PooyaPooya Vahdati Vahdati

Luleå, January 2014

Department of Civil, Enviromental and Natural Resources Engineering Luleå University of Technology SE - 971 87 LULEÅ www.ltu.se/shb

Identification of Soil Parameters in an Embankment Dam by Mathematical Optimization

Pooya Vahdati

Luleå University of Technology Department of Civil, Environmental and Natural Resources Engineering Division of Mining and Geotechnical Engineering Printed by Luleå University of Technology, Graphic Production 2014

ISSN 1402-1757 ISBN 978-91-7439-852-6 (print) ISBN 978-91-7439-853-3 (pdf) Luleå 2014 www.ltu.se

To my parents and brother for their endless love and support

PREFACE

The work presented in this thesis has been carried out at the Division of Mining and Geotechnical engineering at the Department of Civil, Environmental and Natural Resources engineering at Luleå University of Technology. The research has been founded by "Swedish Hydropower Centre - SVC". SVC has been established by the Swedish Energy Agency, Elforsk and Svenska Kraftnät together with Luleå University of Technology, The Royal Institute of Technology, Chalmers University of Technology and Uppsala University.

I would like to express my outmost gratitude to my assistant supervisor Dr. Hans Mattsson for his support, advice and valuable discussions that we have had during this period. In addition, special thanks go to Severine Levasseur who has devoted her time and efforts towards my research, and supervised me along this work successfully.

I am also thankful to Prof. Sven Knutsson for his support and comments which have been extremely valuable to accomplish this work. Furthermore I would like to thank Paul Bonnier from the PLAXIS group for supporting me in this research.

I am grateful to my colleagues at the division for helping create a pleasant working environment; among them I would like to especially thank Jens Johansson for all great discussions during this period.

I

II

ABSTRACT

The finite element method (FEM) has been widely used to analyse earth and rockfill dams. In a finite element analysis a proper constitutive model has to be chosen for each part of the dam in order to simulate the relation between stresses and strains. The zones of an earth and rockfill dam have different functions. Because of that the zones normally consist of various soil types for which the stress/strain response could vary considerably. For each dam zone, suitable values have to be assigned to the parameters included in the constitutive model chosen. In general, laboratory tests and/or field tests of the soil are needed as a basis for this parameter evaluation. However, many dams are old and limited information might be available regarding the soil materials being used in the dam structures. In dams, it is normally very difficult to take up soil samples for testing, especially from the central impervious part, since this might affect the dam performance and the safety of the dam. For dams it would be advantageous if constitutive parameter values could be determined with some non-destructive method.

Inverse analysis provides a possibility to determine the constitutive behaviour of different materials within the dam structure under the condition that the dams have been equipped with various instrumentations, for monitoring dam performance, which record data such as pore pressures, deformations, total stresses and seepage etc. In the method of inverse analysis, two separate parts are included: (1) an optimization method consisting of an error function and a search algorithm and (2) a numerical method to solve the partial differential equations arising in stress-strain analysis of structures. In this study, inverse analysis of a dam case was performed with a commercial finite element program Plaxis and the genetic algorithm was utilized as the search algorithm

III

in the optimization method. The genetic algorithm was chosen due to its robustness and efficiency, particularly since it provides a set of solutions close to the optimum solution instead of one unique answer; a set of solutions is more practical from a geotechnical perspective. In the proposed inverse analysis a finite element model is calibrated automatically by changing the values of the input parameters of the selected constitutive model in different dam zones until the discrepancy between the measured results by dam instrumentations and the corresponding computed results is minimized.

In order to examine the efficiency and robustness of the genetic algorithm, the research was initially focused on a synthetic case study. The synthetic case, a set of model parameters known in advance, is a good test of the mathematical basis used in the optimization, i.e. the objective function and the search algorithm. The Mohr-Coulomb model was chosen for all dam zones, as an initial choice for this research, chiefly because of its simplicity. A very good agreement for the optimization against the synthetic case was obtained. The practical outcome of an inverse analysis clearly depends on the ability of the constitutive models chosen to capture the real soil behaviour in the different dam zones. A proper choice of a constitutive model provides an opportunity to calibrate the finite element model properly. Therefore, in the next step the Hardening soil model, an advanced constitutive model, was chosen for optimization on the dam.

In this part of the research, two cases (A and B) based on different reservoir water levels and number of berms constructed, were analysed. All the data of horizontal displacement were received from exactly the same positions in the geometry as the measurements carried out with the single inclinometer. The results of inverse analyses showed that the Hardening soil model is able to capture better the soil displacements within the dam structure, especially at the crest part, compared to the Mohr-Coulomb model.

Finally, it was concluded that inverse analysis is a practical tool for identifying soil material properties of earth and rockfill dams and provides a non- destructive method for dam engineers to obtain more information about the dams. Moreover, if inverse analysis applications become available in commercial finite element software, it would certainly be a valuable tool for dam engineers assessing dam performance and dam safety.

IV ABBREVIATIONS

Symbol Description Units

c Cohesion in Mohr- Coulomb failure Pa Undrained shear strength Pa

푐D푢 Stiffness matrix N/m E Young’s modulus Pa , Reference Elastic modulus Pa 푟푒푓 Reference modulus of primary loading in 퐸 퐸푟푒푓 Pa drained triaxial test 푟푒푓 Reference modulus of primary loading in 퐸50 Pa drained oedometer test 푟푒푓 Reference modulus of loading/reloading 퐸표푒푑 Pa loading in drained triaxial test 푟푒푓 퐸푢푟g Plastic potential - G Shear modulus Pa First stress invariant Pa

퐼1 Second deviatoric stress invariant Pa 퐽K2 Bulk modulus Pa Earth pressure coefficient at rest -

0 K 퐾 for normally compressed soil - 푛푐 퐾k0x Horizontal0 hydraulic conductivity m/s ky Vertical hydraulic conductivity m/s

V

Elastic swelling modulus Pa

퐾푠 Elasto-plastic compression modulus Pa 퐾m푐 Modulus exponent for stress dependency - Shape factor for Cam clay ellipse/ slope of M - critical state line p Mean stress Pa Mean effective stress Pa ′ 푝 Isotropic pre-consolidation pressure Pa ref′ P푝0 Reference pressure Pa q Deviatoric stress Pa Failure ratio -

푅u푓 Pore pressure Pa α Auxillary model parameter - The unit weight above the phreatic level N/m3 3 훾푢 The unit weight below the phreatic level N/m 훾δ푠 Small increment - ε Normal strain m Elastic strain m 푒 휀 Plastic strain m 푝 휀 Axial strain m 3 휀푎 Volumetric strain 휀푝 Deviatoric strain 푚m 푞 Dummy parameter to describe size of 휀ζ - plastic potential Scalar multiplier -

휆ν Poisson’s ratio - Unload/reload Poisson’s ratio -

휈σ푢푟 Normal stress Pa

VI Normal effective stress Pa ′ , 휎 , Principal effective stresses Pa ′ ′ ′ 휎1 휎τ2 휎3 Shear stress Pa

ϕ Friction angle (°) x Scalar multiplier - ψ Dilatancy angle (°)

VII

VIII CONTENTS

1 INTRODUCTION ...... 1 1.1 Objective of the research ...... 1 2 EARTH AND ROCKFILL DAMS ...... 3 2.1 Introduction ...... 3 2.2 Deformation in embankment dams ...... 4 2.2.1 Abnormal deformation ...... 6 2.3 Finite element analysis of earth and rockfill dams ...... 7 2.4 Instrumentations ...... 8 2.5 Some issues and problems of modelling earth and rockfill dams .... 8 3 INVERSE ANALYSIS ...... 9 3.1 Introduction ...... 9 3.2 Model calibration by inverse analysis ...... 10 3.3 Sensitivity analysis ...... 11 3.4 Case study of an earth and rockfill dam ...... 12 3.5 Finite element model ...... 12 3.6 Constitutive models ...... 15 3.6.1 Introduction ...... 15 3.6.2 Plasticity ...... 16 3.7 Perfect plasticity ...... 21 3.7.1 Mohr-Coulomb model ...... 21 3.8 Cone-Cap models ...... 23 3.8.1 Hardening soil model ...... 25 3.9 Optimization ...... 33 3.9.1 Error function ...... 33 3.9.2 Search algorithm; Genetic algorithm ...... 34 Individuals and populations ...... 35 Initial population, fitness, selection...... 35 Crossover, mutation and mating...... 35 Population size ...... 36

IX Identification of soil parameters in an embankment dam by mathematical optimization

Strategy of running genetic algorithm ...... 37 4 RESULTS ...... 39 5 FUTURE WORK ...... 51 REFERENCES ...... 53

PART II APPENDED PAPERS Paper I Inverse soil parameter identification of an earth and rockfill dam by genetic algorithm optimization Paper II Inverse Mohr-Coulomb soil parameter identification of an earth and rockfill dam by genetic algorithm optimization Paper III Inverse Hardening soil parameter identification of an earth and rockfill dam by genetic algorithm optimization

X Introduction

1 INTRODUCTION

1.1 Objective of the research

In order to model a dam all the material properties of the constitutive models in all the zones within the dam structure needed to be available. But according to the fact that many dams are old and limited information might be available regarding the soil materials being used in the dam structures, as well as restricted recorded instrumentation data from in situ, this purpose is not easy to approach. Moreover, it is normally very difficult to take up soil samples for testing, especially from the central impervious part, since this might affect the dam performance and the safety of the dam. On the other hand, nowadays well- instrumented earth and earth-rockfill embankments provide dam owners and their consultants with a broad database containing pore pressures, deformations, total stresses and seepage etc. A question that arises here is whether any method can be defined for the purpose of identifying the material properties of the dam from recorded data from instrumentations.

Inverse analysis which has been used widely in the engineering and especially in geotechnical field, thanks to the availability of sufficiently fast computer hardware and useful software, has been introduced in this research to provide an opportunity to determine the constitutive behaviour of different materials within the dam structure under the condition that the dam has been equipped with various instrumentations for monitoring dam performance. The advantage of using inverse analysis with e.g. the genetic algorithm as an optimization method is to providing a proper range of material properties of the soil in various zones of the dam out from measured data, without any need for implementing any destructive method. However, engineering judgement is essential to consider for the final evaluation of model parameters.

1 Identification of soil parameters in an embankment dam by mathematical optimization

The objective of this research is to examine if the technique of inverse analysis can be effectively used for identification of constitutive parameter values in an earth and rockfill dam application. If the values of the constitutive parameters can be identified by inverse analysis, reliable finite element simulations of the dam structure can be performed. The results from such simulations can be interpreted and put in relation to dam monitoring data observed to obtain a better overall understanding of the mechanical behaviour of the dam. The dam performance as well as the dam safety can then be better assessed.

2 Earth and rockfill dams

2 EARTH AND ROCKFILL DAMS

2.1 Introduction

A rockfill dam is composed of rockfill, either dumped in lifts or compacted in layers with an impervious face next to water. Rockfill dams are basically embankment dams. High stability and perviousness are their special characteristics. The impervious element in a rockfill dam is provided either by an impervious membrane of manufactured materials or by an earth core. The upstream and downstream slopes of the majority of the rockfill dams with an impervious core of non-earth materials are close to the natural angle of repose, ranging from 1V:1.3H to 1V:1.5H, While slopes of earth-core rockfill dams on the other hand range from 1.5H:1V to 3.25H: 1V (Anon.1993). Due to free drainability and high frictional strength, rockfill dams have a high inherent stability. The earthfill and rockfill dams are often the only real option where the climate is harsh and the geological conditions are unfavourable.

The rockfill dam construction has major advantages in comparison to the other kinds, such as: a wide variety of sites and foundations are suitable, cheap natural construction materials can be found in the local area since even weak rock is also usable and machinery can be used to reduce labour costs so the construction process can be fully mechanized. The rockfill dams with earth cores using the principle of fill-material zoning are mostly economical because all types of locally available rock low-strength (3.5MPa to 17MPa), medium- strength (17MPa to 70MPa) and high-strength rock (70MPa to 200MPa) can be used if the structure is properly zoned; the weaker rock is placed in less critical zones under less stress and hard rock is used where greater strength is required (Anon.1993) .

3 Identification of soil parameters in an embankment dam by mathematical optimization

The accuracy of prediction of rockfill deformation during construction (for both numerical and empirical methods) is dependent on the quality of representation of the stress-strain relationship of the earth and rockfill materials. The stress and strain behaviour in embankment dams, which is governed by loads, particularly: dead weight of soil layers, water reservoir pressure and seismic activity, is predicted from the strength variation and the deformation behaviour of soil under combined stresses. This variation of stress and strain is mainly influenced by: the embankment geometry, the embankment zoning geometry, the strength and compressibility properties of the embankment and foundation soils, the construction technique and speed of construction and the fluctuation in the water reservoir level. The finite element method (FEM) has been widely used to analyse the deformation of earth and rockfill dams regarding changes in stress conditions, mostly during construction and first impoundment (Duncan 1996, Kovacevic 1994, Potts and Zdravković 2001, Muir Wood 2004). In a finite element analysis a proper constitutive model has to be chosen for each part of the dam in order to simulate the relation between stresses and strains. The various zones of an earth and rockfill dam are normally constructed of different soil materials for which the stress/strain response could vary considerably.

2.2 Deformation in embankment dams

Deformations in embankment dams are due to changes in effective stress during construction, impoundment and reservoir fluctuations. A survey of the published case histories indicates that reservoir filling has induced significant movements in many dams, involving formation of potentially dangerous cracks in some well-engineered large dams, and even complete failure of smaller dams (Marsal et al. 1900, Marsal 1959, Marsal et al. 1976, Sherard 1953).

During the first impoundment the vertical and horizontal effective stresses decrease because of buoyancy effects. Decrease in effective stresses leads to reduction in compressive strength of the saturated rockfill material, mainly in the outer surface of the upstream toe. Due to loss of the shear strength of saturated rockfill considerable deformation can occur if the rockfill materials are not well compacted, which is called as wetting deformation, (Nobari and Duncan 1972, Sharma et al. 1979, Terzaghi 1958), see figure (1.a,1.b). This phenomenon on the upstream foundation causes settlement; because the upstream portion of the foundation is wetted first, this settlement is not uniform, see figure (1.c). In the reservoir impoundment, hydrostatic pressure on the upstream face of the core increases with the reservoir level rising, proportionally to the depth of water, see figure (1.d).

4 Earth and rockfill dams

By studying several dam cases Hunter (2003) shows that the dams with poorly compacted rockfill in the downstream shoulder or embankment with core of silty sands to silty gravels had large lateral crest displacements due to reservoir impoundment. He also monitored the displacements of these dams from 5 to 30 years after the first filling. The data shows that for most cases the displacements after the first filling is generally in the range of 35 mm for the upstream shell and from 100 mm to150 mm for the downstream shell. For thin to medium central cores, the water load has a significant effect on the horizontal displacement, close to the centre of the embankment, which leads to increasing lateral stress in the downstream shoulder.

Settlements (vertical displacements) in rockfill dams result from compaction of the fill and foundation under loads from the weight of rock and reservoir pressure. A large part of the total settlement of rockfill happens during construction. The post-construction settlements result from a gradual readjustment of the rock grains structure. The areas at the point of contact between individual rocks grains are small and the stresses are high. Directly after construction, a large deformation may occur when the rock grains crush at the point of contact or slip with respect to each other.

In order to approach a minimum deformation, the dam shell should be filled with strong, unweathered, well graded rock and compacted in layers (Nobari and Duncan 1972). Grading and compressibility are two main geological and physic-mechanical characteristics of the rockfill materials that needed to be considered by dam engineers. The better graded the material, the higher the density of the placed material will be. Dams with well-graded material have higher moduli of compressibility which cause less rock crashing and settlement within the structure. A modulus of compressibility expresses the ability of rockfill to decrease in volume under external load as a result of grain breakage, rearrangement and compaction. Other important factors of rockfill material, which also need to be considered for the design are: strength, permeability, frost and water resistance. For well compacted rockfill materials, the displacements in the long term (more than 10 to 20 years) are typically less than 0.2% of the embankment height, or between 100 to 300mm, while for poorly compacted rockfills, the long term displacements are approximately between 1% to 1.6% of the height of the dam (Hunter 2003).

5 Identification of soil parameters in an embankment dam by mathematical optimization

Figure 1: Effect of reservoir filling on an embankment dam

2.2.1 Abnormal deformation

Identification of abnormal deformation behaviour is made by the comparison to the normal deformation trends in similar embankment types by considering the geometry, material types, placement methods and foundation condition.

These differences are identified by the magnitude of deformations as well as patterns in the direction and rate of deformation measured by surface monitoring of the crest and shoulders. The abnormal behaviour could be specified only for one part of the whole embankment while the overall stability of the embankment is not a major concern. Potentially abnormal deformation patterns include:

• a sudden acceleration in settlement rate at the crest or the upstream slope • plastic displacements in a short period after the first filling • change of direction of displacements.

6 Earth and rockfill dams

In the case study by Hunter (2003), the most abnormal embankment deformations during construction of an earth and rockfill dam are related to the large plastic deformation of the wet part of the core, due to low undrained shear strength, which is caused by high vertical strains and settlements, at the end of the construction. In some cases, a lightly compacted and highly weathered material at the inner rockfill zone could significantly influence large lateral displacement of the core.

2.3 Finite element analysis of earth and rockfill dams

The finite element method was introduced to geotechnical engineering in 1967, when Clough and Woodward (1967) demonstrated its usefulness for analysis of stresses and deformations in earth dams. The finite element method can provide solutions to any level of accuracy and can consider irregular geometry, complex boundary conditions and material non-linearity, which are normally not possible with other analytical methods.

For earth and rockfill dams, the finite element method has been used for computing internal deformation of core and shells, obtaining stress distribution and load transfer within the dam zones and studying the effect of reservoir filling on the stresses and deformations within the core and the other dam zones. Moreover, this technique has been utilized for specifying the location of a foundation cut-off, determination of its stresses and deformation, determining the extent of tensile stresses in a conduit embedded in an earth dam as well as determining stresses and deformations in membrane type rockfill dams. The most important consideration in the application of the finite element method is the method of presenting the complex stress-strain characteristics of soil (generally nonlinear) into a constitutive relationship suitable for incorporation in the finite element model. The stress-strain relationships in soils are path dependent, that is, stresses and strains vary in the manner the loads are applied. Realistic simulation of the stress-strain behaviour of the soil is therefore the factor which mainly influences the accuracy of a finite element analysis of an embankment dam. However, at the end finite element outputs need to have a good and reasonable agreement with the in situ instrumentation data.

7 Identification of soil parameters in an embankment dam by mathematical optimization

2.4 Instrumentations

Dam behaviour during construction and operation, such as: settlement and displacement, stresses obtained, seismic activity, pore pressure generation, seepage and temperature flows are monitored with field instruments in order to be able to analyse the design and construction assumptions.

Field instruments comprise removable devices (installed on the surface of the structure, in section galleries and so forth) and built in instruments (installed within the dam body). Nowadays, most of the built in instruments are remote- reading. Readings are transmitted to receivers in a control room inside the dam or an instrument inspection house.

Recently, a broad database of well-instrumented earth and earth-rockfill embankments, provide dam owners and their consultants with methods that:

• Allow comparison of their structures to similar embankment types in terms of the deformation behaviour during construction, on the first filling and after the first filling. • Broadly define “normal” deformation behaviour and identify potentially “abnormal” deformation behaviour, in terms of magnitude, rate or trend. • Provide some guidance on the trends in deformation behaviour that are potentially indicative of a slightly stable to unstable slope condition, and signs of slope instability.

2.5 Some issues and problems of modelling earth and rockfill dams

Numerical modelling of engineering systems always has a certain amount of uncertainties. These uncertainties are related to the approximations which are made in the conversion of a physical model into a numerical model. In dam engineering, the lack of adequate communication between the design engineers and the engineers supervising the construction of dams makes it difficult to have an accurate judgment of the numerical analysis. For instance, according to the specific geology and topology of the site the construction of an embankment dam might be different from what is described on the design drawings. In addition, the soil properties may be affected by the weather or by the contractor’s method of construction. Therefore modelling of a dam might not be completely matched with the reality and instrumentation data. It is assumed that uncertainties are an unavoidable feature of embankment dam engineering.

8 Inverse analysis

3 INVERSE ANALYSIS

3.1 Introduction

In some geotechnical engineering structures, such as dams, direct identification of the constitutive parameters of soil layers is not always possible. Therefore, computational analysis of such geotechnical structures faces the problem of limited knowledge of mechanical properties of the soil. Inverse analysis provides a possibility to determine the constitutive behavior of different materials within the geotechnical structure. Recently, because of the sufficiently fast and powerful performance of computer hardware and availability of technical software, interest in the application of inverse parameter identification strategies and optimization algorithms for geotechnical modeling in order to make inverse analysis procedures automated, has increased steadily (Swoboda et al. 1999, Gioda and Maier 1980, Tarantola 2005, Calvello and Finno 2004, Finno and Calvello 2005, Ledesma et al. 1996, Lecampion and Constantinescu 2005, Lecampion et al. 2002, Zentar et al. 2001, Levasseur et al. 2010, Yu et al. 2007, Moreira et al. 2013). The inverse analysis technique was introduced to the geotechnical field by Gioda and Sakurai (1987) for the purpose of identification of elastic material properties of in situ rock masses.

In fact, the experience and the judgment of the engineers are a fundamental issue for every inverse analysis. Poor, or even incorrect, results of back analyses are often due to errors in initial engineering decisions, such as material behavior (elastic, elasto-plastic, viscous, etc.), simplification of the numerical model from a physical model and ignorance of the loading history of the soil. In dam applications, inverse analysis provides a possibility to

9 Identification of soil parameters in an embankment dam by mathematical optimization

determine the constitutive behavior of different materials within the dam structure under the condition that the dam has been equipped with various instrumentations for monitoring its performance, which record data such as pore pressures, deformations, total stresses and seepage etc.

3.2 Model calibration by inverse analysis

The model calibration is the process of changing and updating various parts of the model in order to approach a good agreement between measured values (observation data) and computed values. In this research, finite element models are calibrated based on an iterative automatic procedure, correcting the trial values of the selected constitutive model parameters until the discrepancy between computed output and observed data is minimized, see figure (2). The minimization procedure is guided by an error function that evaluates the discrepancy between the results from simulation and the experimental data.

Initially, it is essential to select the optimization parameters and to decide on reasonable upper and lower limits of their values. The relative importance of the constitutive parameters is estimated by parameter sensitivity analysis, performed during a first approach to numerically simulating the observation data. After defining the optimization parameters for calibration, the values of other constitutive parameters will be fixed with the aid of previous studies. This choice of optimization parameters must be made with care since it influences the outcome of the analysis and it requires a lot of engineering experience. The values of constitutive parameters could be estimated by conventional means, e.g. using laboratory tests and field tests. However, in this study, due to the application of a genetic algorithm as search strategy in optimization initial guesses for the set of model parameters are generated randomly within the search space, i.e. the upper and the lower limits of the parameter values. After defining the optimization parameters, a numerical simulation of the problem is performed and the results are compared with the field observations by the error function. The error function quantifies the deviation between computed results and observations. The updating (optimizing) of the optimization parameter values is continued till the discrepancy between computed results and observation is minimized, or the convergence criteria are satisfied.

In the following sections of this chapter, the case study and suggested constitutive models of this research will be introduced. The model will be completed by introducing the optimization method chosen for this research at the end of this chapter.

10 Inverse analysis

Input optimization parameter

Numerical model

(FE model)

Search algorithm Numerical results (optimization) (FE results)

Evaluate the Observation error function measurements

No Yes Model Optimized? End

Figure 2: Schematic flowchart of inverse analysis procedure

3.3 Sensitivity analysis

A sensitivity analysis in this context, is the prediction of the relative sensitivity of the values of the constitutive model parameters in numerical modelling of the values of the same quantities as are measured in the real system; here inclinometer data was measured in the dam in the case study. Initially, the constitutive model needs to be defined in order to specify the total number of model parameters. Then, for each parameter of the constitutive model a sensitivity study is performed. The studied parameters are given different values within a given range, while all the other parameters of the model have fixed values, and for each value of the studied parameter a finite element

11 Identification of soil parameters in an embankment dam by mathematical optimization

computation is performed. The results of these studies are compared to each other in order to define the most sensitive parameters for optimization against observation data in the error function.

3.4 Case study of an earth and rockfill dam

The earth and rockfill dam used for this case study is 45m high, with a dam crest of 6.5m, see figure (3). The upstream slope is inclining 1V:1.71H and the downstream slope is inclining 1V:1.85H. The impervious core with moraine material is surrounded by fine and coarse filters and supporting fill of blasted rock. The dam was strengthened in two stages with rockfill berms on the downstream part. The first stabilizing berm of rockfill was constructed in 1990 and the second berm was constructed three years later in 1993. The dam is well equipped with instrumentations and monitoring programs including deformation measurements within the dam structure. Horizontal displacements are measured automatically by an inclinometer installed along the downstream fine filter, from the crest into the rock foundation. Since the inclinometer data from the ten first meters from the bottom of the core seem not to be trustworthy, only data above this level are considered. The level of water in the reservoir varied from +430m to +440m annually.

Figure 3: Cross section of the dam

3.5 Finite element model

In soil mechanics, complex problems are usually solved by using the powerful and reliable computational capabilities of the finite element method coupled with proper constitutive models. Geotechnical engineers need to have a sufficient understanding of both geotechnical and finite element theory in order to perform accurate numerical analyses.

12 Inverse analysis

The Plaxis finite element program (2011) was applied to the numerical analyses in this study. Plaxis is developed for analysis of deformation and stability in geotechnical engineering structures. The finite element model of the cross section of the dam is shown in figure (4). Since the dam is a long structure, plane strain condition was assumed. Some smaller simplifications compared to the real dam cross section in figure (3) were made in the finite element model in figure (4), e.g. the upstream toe berm was neglected due to similarity in its material properties with rockfill. Horizontal filters were omitted as well, since consolidation time is not analyzed. These simplifications are considered to be of no practical importance in this study. For the analysis to be accurate enough, a sufficiently large part of the cross sectional geometry was considered in the finite element model. The option standard fixities in Plaxis were selected for generation of boundary conditions to the geometry model, for the outermost vertical left and right foundation boundary horizontal displacements were prescribed to zero and for the lowest horizontal boundary in the foundation both horizontal and vertical displacements were prescribed to zero. The element type was chosen as 15-node triangular elements. The coarseness of the mesh was selected so that a sufficient accuracy of the computations was obtained for a minimum of computation time spent. Several different meshes were tested and the one illustrated in figure (4) was chosen. The dam embankment in the model in figure (4) was constructed stepwise in five stages in order to generate a proper initial effective stress field (Duncan 1996, Reséndiz and Romo. 1972, Clough and Woodward 1967). In the analyses both the Mohr-Coulomb model and the Hardening soil model were utilized. The excess pore pressures in the core were dissipated by consolidation after construction of the dam. The parameter values of γ u ,γ s , φ', kx and ky for all soil layers in the dam were provided by the dam owner. Appropriate values of ν' and c' were chosen based on advices in the reference (Bowles 1988). All the dilatancy angles were calculated from the same empirical relation, commonly accepted for dense sand: ψ' = φ'-30° as proposed by Plaxis (2011).

13 Identification of soil parameters in an embankment dam by mathematical optimization

Figure 4: Finite element model of the cross section of the dam

In the finite element analyses executed in this thesis, simulated horizontal deformations received in exactly the same positions in the geometry as the measurements carried out with the single inclinometer in figure (3), were compared to the corresponding inclinometer measurements in the expression for the error function. In the finite element analyses, all displacements were reset to zero before impoundment. In this study the deformation comparison was performed for two cases (A and B) based on different reservoir water levels and berm constructions.

Case A is related to the time that the first stabilizing berm was constructed in 1990 and the reservoir water was at the minimum level of +430 m, figure (5).

Case B is when the reservoir water level reached its maximum level +440m and the second berm was just finished. The second berm was rapidly constructed right after the water level reached +440m, figure (5).

Case A Case B Figure 5: Cross sections of cases A and B

14 Inverse analysis

3.6 Constitutive models

3.6.1 Introduction

Soil behavior is very complicated and can vary considerably under different loading conditions. Real soil is normally a highly nonlinear material, with both strength and stiffness properties depending on stress and strain levels. Fundamentally, soil behaves as an elastic-plastic material which means that the deformations are inelastic since upon load removal, unloading follows a completely different path from the loading path. No constitutive model can completely describe the complex behavior of soils under all loading situations. Therefore, constitutive models have limitations for their applicability. All models have certain advantages and limitations which highly depend on their particular application. For instance, Hook’s law (based on linear elastic theory) has been successful in describing the general behavior of soil under short working load situations, while this model is not able to capture the plastic behavior of soil and it fails to predict properly the compressibility and strength of soil near ultimate strength conditions. Often hyperbolic soil models, such as the Duncan-Chang model (Duncan and Chang 1970, Duncan 1981) have been used in geotechnical problems to describe the nonlinear stress-strain behavior of soils. But since this model is a purely hypo-elastic model, it cannot distinguish between loading and reloading, see figure (6). Even advanced constitutive models such as the Hardening soil model have limitations, it is e.g. not able to capture the strain softening behavior of some soils types under special drainage type (drained triaxial test on dense sand and over-consolidated clay, undrained triaxial test on normally consolidated clay and loose sand (Chen and Mizuno 1990). However this model can perfectly follow the elasto- plastic behavior of soil up to the peak stress point.

Figure 6: Duncan-Chang model, presenting loading and unloading path

15 Identification of soil parameters in an embankment dam by mathematical optimization

The selection of a proper constitutive model depends on the required accuracy for the analysis, the type of loading conditions and undrained or drained behavior etc. For soils, many times a perfect- plastic model is an appropriate simplification, although more complex stress-strain relations of soil may be approached by the more practiced strain hardening theory of plasticity.

3.6.2 Plasticity The application of plasticity theory to geomaterials has been developing fast in recent years. Although more progress is expected, the plasticity theory for geomaterial may have approached a good level of development. Nowadays, no one would doubt the importance of plasticity theory in geotechnical engineering. A review of the early development can be found in Von Mises (1928), Melan (1938), Drucker (1950), Prager (1955), Hill (1998), Naghdi (1960), Yu (2006). The recent developments of plasticity theory for geomaterials are based on achievements in metal plasticity. However, unlike metal materials, volume changes during loading are one of the key factors in modelling plastic behaviour of soil materials.

The plasticity theory can be explained by introducing flow theory or incremental theory of plasticity. The flow theory is an appropriate extension of elastic stress-strain relations to the plastic criteria in which the total strain is the combination of reversible elastic strain and irreversible plastic strain. Elastic strain increments can be described within the framework of Hooke’s law on incremental form, see section 3.7.1. On the other hand, in order to estimate the plastic strain increments or formulate the incremental theory of plasticity three main assumptions need to be defined: (1) yield criterion (2) hardening rules (3) flow rule. The plastic strains are the combination of two independent parts, the shear plastic strain, , and the volumetric plastic strain, , a concept which was introduced by Roscoe푝 and Burland (1968) 푝 휀푞 휀푝 = + (1) 푝 푃 or to simplify the notations 훿휀푝 훿휀푞 훿휀푝 = + (2) 푝 푝 휀푝̇ 휀푞̇ 휀푝̇ Yield criterion

The concept of a yield locus or a yield surface is one of the most important parts of the theory of plasticity. A condition that defines the limit of elasticity and the beginning of plastic deformation under any stress condition is known

16 Inverse analysis as the yield criterion, see figure (7) the yield criterion is represented by a surface in stress space (a curve in the p-q plane), which is called yield surface. When the stress state is located within the yield surface, material behaves as elastic. Once the stress state is on the yield surface, plastic deformation will occur. Generally, the yield surface expands or contracts depending on previous plastic deformations and loading history. However, for perfectly plastic material, the yield surface is fixed in stress space; therefore plastic deformation happens only when the stress path reaches the yield surface. Plastic strain occurs as long as the stress state either lies or travels along the yield surface. When the yield surface is expanding in size, the material is said to be hardening. On the other hand, when the yield surface is contracting in size, then the material is said to be softening. The change in yield surface size is a function of plastic strain rates, normally a function of the volumetric plastic strain rate alone (Schofield and Worth 1968, Yu 1998, Wood 1990) or a function of both푝 volumetric and shear plastic strain rates (Wilde 1977, 휀푝̇ Yu et al. 2005). 푝 푝 휀푝̇ 휀푞̇

Figure 7: yield surfaces in p-q plane

Combinations of mean effective stress, , and deviator stress, q can be applied to samples of soil in the triaxial test to′ obtain information about the yield surface for the soil. The invariants, 푝 and q, used by Roscoe and Burland (1968), defined in terms of all principal ′stresses 푝 1 = ( + + ) (3) 3 1 ′ ′ ′ ′ = [( 푝 ) + (휎1 휎2 휎)3 + ( ) ] (4) 2 1 ′ ′ 2 ′ ′ 2 ′ ′ 2 2 푞 휎1 − 휎2 휎2 − 휎3 휎3 − 휎1 √ 17 Identification of soil parameters in an embankment dam by mathematical optimization

The isotropic pre-consolidation pressure, , is a parameter which controls the expansion or contraction of the yield surface′ i.e. the hardening parameters. Thus, the yield surface can be expressed by푝0 the following relation

= ( , , ) = 0 (5) ′ ′ ′ Hardening rule 푓 푝 푞 푝0 The hardening rule is a law which describes the change in size of the yield surface due to the process of incremental plastic deformation. There are a number of hardening rules that have been proposed in the literature, but the choice of a specific hardening rule depends on its simplicity and its ability to represent the hardening behaviour of the soil. Isotropic hardening, kinematic hardening and mixed hardening are three types of hardening rules which have been introduced foremost in metal plasticity (Chen 2007). The isotropic hardening model is the simplest work hardening model among these three models. It assumes that the yield surface expands uniformly about a centre without changes in shape as plastic flow occurs, see figure (8). In the figure (7), the hardening rule relates the change in size of the yield locus; change in , to the increments of the plastic volumetric strain and the plastic shear strain,′ according to the incremental relation 푝0

= + ′ ′ (6) ′ 휕푝0 푝 휕푝0 푝 훿푝0 푝 훿휀푝 푝 훿휀푞 휕휀푝 휕휀푞 An example of an isotropic strain hardening model is the Hardening soil model described in 3.8.1 (Chen and Mizuno 1990).

18 Inverse analysis

Figure 8: Isotropic hardening

Flow rule

As mentioned before, yielding is related to the occurrence of plastic volumetric strain and plastic shear strain. Yielding can happen under many combinations of stresses in the loading history of a soil. Yielding can be presented by the plastic strain vectors, see figure (9).

Figure 9: Plastic strain increment vectors normal to the plastic potential curves (Muir Wood, D)

19 Identification of soil parameters in an embankment dam by mathematical optimization

A line, like AB, can be drawn orthogonally to each plastic strain increment vector, like YS, passing through each yield point, like Y. As more yielding data become available in soil tests the lines that are orthogonal to the plastic strain vectors in each yield point, can be joined and make curves. These curves are called plastic potentials. The direction of the outward normal to the plastic potential surface defines the relative magnitude of the plastic shear strain increment (change in shape) and the plastic volumetric strain increment (change in size). The plastic potential expression in soil mechanics can be formulated as

( , ,ζ)=0 (7) ′ ′ 푔 푝 푞 where ζ is a parameter controlling the size of the plastic potential, the same as purpose for a yield locus. The plastic strain increment vector relates to the gradient vector′ of the plastic potential curve at a current effective stress state, in the form푝0 of the flow rule

= (8) 푝 휕푔 훿휀푝 휆 ′ = 휕푝 (9) 푝 휕푔 훿휀푞 휆 휕푞 which 0 is a scalar multiplier, which controls the magnitude of plastic deformations. The components of the gradient vector and are the 휆 ≥ 휕푔 휕푔 parameters which control the direction of the plastic deformations.′ Notice that 휕푝 휕푞 the plastic strain increment vector becomes normal to the plastic potential curve since the gradient vector is normal to the plastic potential curve. Only partial derivatives of the plastic potential curve are utilized in plasticity, which is the reason for naming ( , ,ζ) the potential function. With knowledge of the plastic potentials, the mechanism′ ′ of plastic deformation is defined. When it comes to formulating constitutive푔 푝 푞 models for soil, it is really convenient if, the shape of the yield loci and the plastic potential can be assumed to be the same. Then only one function has to be defined. In this case = the material is said to follow a law of associated flow. Otherwise, if it is called non- associated flow. In figure (10), sketches of associated and푓 non푔-associated flow are presented. 푓 ≠ 푔

20 Inverse analysis

Figure 10: Associated and non-associated flow rules

3.7 Perfect plasticity

The concept of perfect plasticity has been used for quite some time for geotechnical problems, for example stability involving earth pressure and retaining walls by Coulomb (1776) and Rankine (1857). In these analyses, the maximum load that a structure can take before collapsing was calculated based on rigid-perfectly plastic soil behaviour. In many engineering applications, soil and rock can be modelled as an elastic-perfectly plastic solid as a simple first approximation.

3.7.1 Mohr-Coulomb model The Mohr-Coulomb criterion is the best known failure criterion in soil mechanics. This criterion states that failure happens when the shear stress τ and the effective normal stress acting on any element in the material satisfy the linear equation ′ 휎 | | = + tan (10) ′ 휏 푐 휎 휙 where c and are the cohesion and the internal angle of friction, respectively. In terms of the principal effective stresses, Coulomb’s yield criterion, see figure (11), can휙 be expressed by the following equations 1 1 = ( ) ( + ) sin cos = 0 (11) 2 2 13 ′ ′ ′ ′ 푓 휎1 − 휎3 − 휎1 휎3 휙 − 푐 휙

21 Identification of soil parameters in an embankment dam by mathematical optimization

1 1 = ( ) ( + ) sin cos = 0 (12) 2 2 12 ′ ′ ′ ′ 푓 휎1 − 휎2 −= 휎 (triaxial1 휎2 compression)휙 − 푐 휙 ′ ′ ′ This failure criterion 휎can1 ≥ be휎2 re-written휎3 in terms of deviatoric stress and mean effective stress which is a more common expression for defining failure behavior in advanced soil models

6 sin = ( + cot ) (13) 3 sin 1 휙 ′ In the Mohr-Coulomb푓 model푞 − the yield 푝surface푐 is 휙fixed, therefore the yield surface and the failure surface become− 휙 the same, see figure (11). Shield (1955) shows that Coulomb’s failure criterion is an irregular hexagonal pyramid in a three dimensional principal stress space; see figure (11). The Mohr-Coulomb surface shows singularities along the corner edges, which cause some difficulties in numerical analysis since the yield functions are non- differentiable at the corners. This problem can be solved by Kioters rule (Koiter 1960) . In addition to that, this model neglects the influence of intermediate effective principal stress on the shear strength. This model also in its basic form, includes neither stress-dependency nor stress-path dependency of stiffness. In the Mohr-Coulomb models it is assumed that the material behaves elastically until the failure surface is approached. However, in reality, plastic deformations begin well before the failure occurs. But still this model has been used because of its simplicity and capacity to obtain reasonable solutions to important practical problems in geotechnical engineering.

Figure 11: Mohr Coulomb’s failure surface

22 Inverse analysis

The stress-strain relation is expressed based on Hooke’s law in terms of Young’s modulus and Poisson’s ratio. Alternatively, it can be expressed in terms of bulk modulus K and shear modulus G with the transformations

E K = (14) 2(1-2ν) E G = (15) 2(1+ν)

In Plaxis, besides utilizing this model based on a linear elastic relation, an option exists that specifies a stiffness that varies linearly with depth. By specifying an increment of stiffness , the value of the elastic modulus increases linearly regarding the depth of the soil. The relation of incremental 푖푛푐 stiffness is 퐸

= + ( ) (16)

퐸 퐸푟푒푓 푦푟푒푓 − 푦 퐸푖푛푐 where y is the position of a selected point with respect to the position of the reference point and is the elastic modulus related to the reference point. So, the alternative elastic incremental parameters can be changed to 푦푟푒푓 퐸푟푒푓 = (17) 2(1-2ν) 퐸푖푛푐 퐾푖푛푐 = (18) 2(1+ν) 퐸푖푛푐 퐺푖푛푐 3.8 Cone-Cap models

The cap types of strain-hardening models are models which are capable of treating the conditions of unloading, stress path-dependency, dilatancy and the effect of intermediate effective principal stress. Drucker et al. (1957) proposed that soil might be modelled as an elastic-plastic strain hardening material, see figure (12). They introduced a spherical end cap to close the former Drucker- Prager cone (Drucker and Prager 1952), in order to control the plastic volumetric strain of the soil. As the soil strain hardens (effective stresses increase) plastic deformations are generated and both the cap and the cone expand.

23 Identification of soil parameters in an embankment dam by mathematical optimization

Figure 12: Drucker-Prager type of strain-hardening cone-cap model

The mathematical formulation of the cap part satisfies the Drucker-Prager uniqueness of stress-strain relations. This model consists of both a failure surface for perfectly plastic response

= ( , ) = 0 (19)

=푓푓 + 퐼1 퐽+2 = 3 (20) ′ ′ ′ ′ 1 1 2 3 퐼 휎 3휎 = 휎 푝 (21) where is the first stress invariant,2 the second deviatoric stress invariant. � 퐽 푞 And an elliptic strain-hardening cap 퐼1 퐽2 = ( , , ) = 0 (22)

푓푐 퐼1 퐽2 푥 where x is the hardening parameter for the cap. Therefore, three types of material response might be represented: elastic, perfectly plastic and strain hardening, under the assumption of an associated flow rule. Changes in the size of the cap is due to changes in volumetric strain through the hardening parameter x, see figure (13).

24 Inverse analysis

Figure 13: Elliptic cap model in -space

In general, cone-cap models are applicable for complicated퐼1 − �퐽2 situations such as nonlinear loading and unloading, stress path dependency and dilantancy. Hardening plasticity theories to describe the stress-strain behaviour of soil have been highly developed recently in soil mechanics, especially for cyclic loading problems.

3.8.1 Hardening soil model The Hardening soil model is a cone-cap model (Schanz et al. 1999). The plastic strain behaviour on the cone is described by means of a shear yield surface following a non-associated flow rule based on Rowe’s stress-dilatancy theory (1962). The plastic potential is defined to assure a hyperbolic stress/strain response for a triaxial compression loading. The cap yield surface is controlled by the volumetric plastic strain. On this cap, an associated flow rule is assumed. This yield surface is defined in order to close the elastic region in the -axis direction, see figure (14). The failure on the cone part is defined based on′ Mohr-Coulomb’s failure criterion which is characterized by the strength parameters푝 c, φ and the dilatancy angle ψ. The shape of the cone yield surface is “flattened out” when it approaches the critical failure line. The Hardening soil model uses a hyperbolic stress-strain curve instead of a bi-linear relation as in the Mohr-Coulomb model. Since in reality the stiffness of soils depends on the stress level, the Hardening soil model is formulated to capture this stress level dependency.

25 Identification of soil parameters in an embankment dam by mathematical optimization

Figure 14: Yield surface of hardening soil model in p-q plane

The experimental basis for the cone yield surface is taken from two different triaxial compression test series which were reported by Tatsuoka and Ishihara (1974). In the testing procedure for the first test series the deviatoric stress q was initially increased, thereafter reduced and then increased again. This was performed for different effective mean stresses . From this procedure of deviatoric loading and unloading, complete successive′ yield loci were found, see figure (15). The second tests series involved loading푝 various virgin samples for different loading paths, including undrained tests. From the results of these tests, it was concluded that the shear strain is dependent on the final stress state only and not on the shape of the stress path. In other words, the yield locus can be approximated to curves of constant shear strain, at least for primary loading, see figure (15). It was found by Vermeer (1978) that for sand and sandy materials the difference between plastic shear strain and total shear strain is small. Therefore, in this model we consider the total shear strain as the plastic shear strain.

26 Inverse analysis

Figure 15: Data from loose samples after Tatsuoka and Ishiara: Shear yield loci and shear strain contours

Two yield surfaces were introduced from the triaxial test results, based on plastic shear strain

( ) 2( ) = ( ) (23) (′ ′ ) ′ ′ 푞푎 휎1 − 휎2 휎1 − 휎2 푝 푝 푓12 ′ ′ − − 휀1 − 휀2 퐸50 푞푎 (− 휎1 − )휎2 2( 퐸푢푟 ) = ( ) (24) (′ ′ ) ′ ′ 푞푎 휎1 − 휎3 휎1 − 휎3 푝 푝 푓13 ′ ′ − − 휀1 − 휀3 퐸50 푞푎 − 휎1 − 휎3 퐸푢푟 where is the confining stress dependent stiffness modulus for primary loading and is elastic unloading/reloading parameter, see table (1). The 50 ultimate퐸 deviatoric stress, , is defined from the Mohr-Coulomb failure 푢푟 criterion, which퐸 includes the strength parameters c and 푞푓 휙 6 sin = ( + cot ) (25) 3 sin ∅ ′ 푞푓 푝 푐 ∅ as soon as q = , the failure− criterion∅ is reached and perfectly plastic yielding occurs. 푞푓

27 Identification of soil parameters in an embankment dam by mathematical optimization

The plastic volumetric strain which is measured in isotropic compression also needs to be defined. Therefore, the second yield surface, the cap surface, see figure (14), must be introduced in order to close the elastic region in the direction of the -axis ′ 푝 = + ( + ) ( + ) (26) 2 푞� ′ 2 ′ 2 where M is an auxiliary푓푐 model2 parameter푝 푎 −that푝 푐relates푎 to and a is defined as 푀 cot . Furthermore 푁퐶 퐾0 푐 ∅ = + ( 1) (27) ′ ′ ′ where 푞� 휎1 훼 − 휎2 − 훼 휎3

= (regular yield surface) (28) 3+sin ∅ in the special case of훼 triaxial3−sin ∅compression = = ( ) and for triaxial extension = ( ). As mentioned before, for ′ the cap′ an associated 1 3 flow rule is assumed′ with′ the plastic potential푞� 푞= 휎 − 휎 푞� 훼 휎1 − 휎3 푔푐 푓푐 = + = + 2 (29) ′ ′ 휕푔푐 ′ here H is the 푝hardening푐 푝푐0 휆modulus퐻 ′ which푝푐0 describes퐻 휆 푝 the ratio between compression and swelling stiffness휕 in휎 isotropic loading

= (30) 퐾푐 where is the elastic 퐻swelling modulus퐾푠 and is the elasto-plastic 퐾푠 − 퐾푐 compression modulus. The plastic multiplier refering to the cap is determined 푠 푐 based on퐾 the following equation 퐾 휆 + = (31) 2( + ) ′ + 푚 ′+ 퐻 푝푐 푎 푝̇푐 When the stress 휆point is ′situated� on푟푒푓 the intersection� 푟푒푓 between the cone and the 푝 푎 휎 푎 휎 푎 cap, there are various possibilities for loading and unloading. Figure (16) shows all these possibilities in the -q plane. ′ 푝

28 Inverse analysis

From the linear elastic law

= = ( ) = (32) 푒 푝 푒 푝 휎̇ 퐃휀̇ 퐃 휀̇ − 휀̇ 휎̇ − 퐃휀̇ and with association of the flow rule it becomes

= = (33) 푒 휕푔 푒 휕푔 휎̇ 휎̇ − 퐃 휆 휎̇ − 휆퐃 휕휎 휕휎 If loading occurs in such a way that several yield surfaces become active, then the total plastic strain increment is assumed to be the sum of plastic strain increments from all the surfaces 푝 휀̇ = + + (34) 푝 푝 푝 푝 휀̇ 휀푓̇ 휀푐̇ 휀푚̇ the subscript f is for defining the friction hardening, c is for the cap hardening and m is for the Mohr-Coulomb criterion. Moreover, by considering the flow rule, it becomes

= (35) 푓 푐 푚 푒 휕푔 휕푔 휕푔 휎̇ 휎̇ − 휆푓퐃 − 휆푐퐃 − 휆푚퐃 휕휎 휕휎 휕휎 since the stiffness matrix D is constant, then

= (36) 푒 휕푔푓 휕푔푐 휕푔푚 The non-negative휎 multipliers휎 − 휆푓퐃 λ have− to휆푐 퐃be determined− 휆푚퐃 from 휕휎 휕휎 휕휎 , , 0 ; , , 0 (37) If we consider the initial stress point from point “O” then the “OB” path 휆푓 휆푐 휆푚 ≥ 푓푓 푓푐 푓푚 ≤ represents friction hardening only, since the cap size remains the same. The “OF” path is on the other hand shows cap expansion without change in frictional hardening, therefore, this path shows pure cap hardening. The combination of both cap and frictional hardening can be seen in the path “OE”. The points C and D are located on the Mohr-Coulomb failure line; therefore the path “OC” represents frictional hardening until the Mohr-Coulomb

29 Identification of soil parameters in an embankment dam by mathematical optimization

criterion is reached. And the path “OD” represents frictional hardening until the Mohr-Coulomb criterion is reached as well as cap hardening. Since the Mohr-Coulomb criterion limits the frictional hardening (the side of the cone) we never have both 0 and 0. Therefore, at most two yield functions are active. 휆푓 ≥ 휆푚 ≥

Figure 16: Yield hardening under different stress paths

Figure (17) shows all the plastic points of the Hardening soil model for this research. These points represent the final stress state in the earth and rockfill dam structure. Figure (18a) shows the location of plastic points on the yield loci of this model. The deviatoric stress and mean effective stress change according to the water load (lateral stress) and weight of the soil (vertical stress). The final stress state can be located on any friction or cap hardening loci close to the mean effective stress axe or close to the failure line depending on the stress condition of that point, see figure (18b). For instance, the friction hardening points in figure (17) are along the core from the top to the bottom located on different friction hardening yield loci, see figure (18b).

The plastic points are situated both on the cone and the cap of the Hardening soil model. It’s interesting to notice that all mathematical parts of the model are active during the numerical simulation of the dam.

30 Inverse analysis

MC-failure points Cap and friction hardening points

Cap hardening points Friction hardening points

Tension cut-off points

Figure 17: Plastic points on different yield loci

(a) (b) Figure 18: Plastic points on Hardening soil model a) yield loci b) Multi hardened yield loci

31 Identification of soil parameters in an embankment dam by mathematical optimization

Table 1: Parameters of the Hardening soil model

Parameter Definition Relation Slope of the failure line in ϕ Friction angle stress space plane ′ ′ Failure ratio τ − 휎 = 푞푓 푅ψ푓 Dilatancy angle ψ 푅=푓 ϕ 푞-푎30° ′ ′ ′ Reference secant stiffness for ' m ref ref c cot ϕ - E primary loading in a drained E50=E 50 50 c' cot ϕ ′-Pref′ triaxial test 휎3 � ′ � ' m ref Reference tangent stiffness for ref c cot ϕ - E Eoed=E oed primary oedometer loading oed c' cot ϕ ′-Pref′ 휎1 � ′ � Reference stiffness for ' m ref ref c cot ϕ - Eur unloading\reloading in a Eur=Eur c' cot ϕ ′-Pref′ drained triaxial test 휎3 � ′ � Slope of trend line in Power for stress-level m log ref′ - log E50 space dependency of stiffness P 3 휎 plane � � Poisson's ratio for ν 0.15-0.30 ur loading\unloading

Earth pressure coefficient at K0 K0= ′ rest 휎3 ′ Earth pressure coefficient value 휎1 Knc Knc= 1- sin ϕ 0 for normally consolidated soil 0 ′ Note: is the ultimate deviatoric stress, is the asymptotic deviatoric stress 푓 푎 푞 푞

32 Inverse analysis

3.9 Optimization

The aim of optimization is to approach the optimum solution among other solutions for a given problem as a matter of efficiency. The problems are often models of the physical reality. The applicability of optimization can be found in almost every real world system in which numerical information is processed (engineering, mathematics, economics, etc.). For any kind of optimization an error function needs to be defined in order to qualify the deviation.

Many different optimization methods have been proposed in the literature. The best choice of optimization method depends strongly on the type of optimization problem that should be solved. A very important factor in the choice is the smoothness properties of the error function surface. For instance, gradient methods (e.g. Calvello and Finno 2002, 2004, Lecampion et al. 2002, Ledesma et al. 1996) are commonly regarded as efficient, but they use derivatives of the error function and are therefore only practical when the error function is continuous and differentiable. Direct search methods, like e.g. the simplex method by Nelder and Mead (1965), in which the search strategy is based only on values of the error function itself, can be easily trapped in local optima in the search domain.

As the search method for finding the minimum value of the error function for the dam applications in this work, the genetic algorithm was chosen. The variables to optimize consist of selected parameters included in constitutive models utilized for the finite element simulations. Notice that the magnitude of the error function is thereby influenced by the values of the optimization variables. The optimal values of the chosen parameters are progressively approached through iteration by minimizing the error function; that is the optimum solution.

3.9.1 Error function

In order to define the difference between the measured values and finite element computed values, it is essential to define a proper error function. In this research, following Levasseur et al. (2008) a scalar error

function Ferr based on the least-square method was introduced

1 N 2 2 1 Ue - Un i i � (38) Ferr = N ΔU 2 i=1 � i � � � �

33 Identification of soil parameters in an embankment dam by mathematical optimization

that determines the difference between experimentally measured values

Ue i and numerically computed values Uni, from the N measuring points. The term1/ΔUi gives the weight of the discrepancy between Ue i and Uni.The parameter ΔUi relates to the numerical and experimental uncertainties at the measurement point i; as

Ui = ε + αUe i (39) where the parameter ε is an absolute훥 error of measurements and the parameter α is a dimensionless relative error of measurements.

3.9.2 Search algorithm; Genetic algorithm

It is generally accepted that evolution occurs by natural selection, within a given environment, more successful creatures survive and reproduce, whereas the less well adapted or weaker decline. We could consider evolution as an effectively self-optimizing process. The genetic algorithms are search procedures based on the mechanics of natural selection and natural genetics. Genetic algorithms are based on the concepts of biological evolution, applying the theory of survival of the fittest to produce successively better approximation to the solution. In the real world, an organism’s characteristics are encoded in its DNA. Genetic algorithms, by following the same pattern, store the characteristics of an artificial organism in a long string of bits (genotype).

Genetic algorithms were developed originally in the field of artificial intelligence by john Holland in the 1970s at the University of Michigan and popularized as a universal optimization algorithm. The Genetic algorithms are computationally and numerically simple and powerful in their search for improvement. Furthermore, they are not restricted by smoothness properties of the error function surface, like continuity and existence of derivatives.

The genetic algorithms are global optimization techniques that are based on randomized operators such as: selection, crossover and mutation. The search methods can approach a set of solutions close to the optimum solution without being trapped into local optima in the given search domain. Genetic algorithms have been applied in many fields and practical applications in engineering systems. Stoffa and Sen (1991) have used the approach for fitting synthetic body waveforms for 1-D depth seismic profiles; Kennett and Sambridge (1994) and Gallagher et al. (1994) have identified earthquake hypocenter locations;

34 Inverse analysis

Sundaram and Venkatasubramanian (1998) used a genetic algorithm for studying large scale polymer design problems. This method has been utilized in some geotechnical applications like excavation and pressuremeter test studies; see e.g. (Levasseur et al. 2010, Levasseur et al. 2008, Levasseur et al. 2009, Rechea et al. 2008), Papon et al. (2012). The genetic algorithm consists of several operators which will be discussed more closely in the following chapters.

Individuals and populations

Individuals are the single solutions. An individual is a genetic algorithm chromosome which has raw genetic information. Each gene is encoded into a part of the bit string in order to be passed to the numerical model. Thus, each optimization parameter is encoded to a form of gene. The number of optimization parameters Np define the dimension of the search space limited by the upper and the lower bound constraints of each parameter. The Population, Ni , is the set of individuals (solutions) involved in the search domain. Initial population, fitness, selection

The initial population is the number of chromosomes, each randomly chosen to fill the defined population in the Np domain. Initially, these random solutions of the initial population need to be evaluated. Based on the fitness value of the individuals, the population is sorted in ascending or descending order. In the genetic algorithm the fitness is the value of the error function for its associated individual. The error function provides a measure of how individuals have performed in the search area. With the aim of collecting the best individuals, only Ni/3 of them are retained, selected, for the next population, which are called parents.

Crossover, mutation and mating

The crossover’s purpose is to create new solutions (children) by combining the selected good temporary solutions (parents). The children get some parts of both parents’ genetic material. The most common form of mating includes two parents that produce two children (offspring). The position of crossover points is randomly chosen and the value of each binary string is exchanged for with another one after this position point, see figure (3). Each parent copies part of the gene to each child, with the selection of the genes being chosen independently. This process is repeated until 2Ni/3 children are created. Now the new population, consisting of parents and children is completed.

35 Identification of soil parameters in an embankment dam by mathematical optimization

Parent #1 111001 100100 111001 011011 Offspring #1

Parent #2 001110 011011 001110 100100 Offspring #2

Figure 19: Reproduction of offspring (children) from the pair of parents

Mutation is the variation of a randomly selected bit in the parameter code based on a specific probability. Goldberg (1989) recommends low nonzero mutation probability, e.g. 0.001, in order to allow diversity to occur within the population. The role of mutation in genetic algorithms is to search unexplored areas of the search domain to prevent premature convergence of the genetic algorithms to a local optimum. Selecting a proper value for mutation probability is not an easy task. Large values change the genetic algorithm into a decently random search algorithm and low values omit the effect of mutation in the process. The typical value for mutation probability is in the range of 0.005- 0.05. The procedure of selection, crossing, mating and reproduction is called breeding or mating.

Selected gene before mutation 001011010110 Selected gene after mutation 001010010110

Figure 20: Mutation of an individual

Population size

The size of the population must initially be large enough to achieve a good convergence of the algorithm. To avoid a prohibitive CPU time, it is then necessary to reduce dynamically the size of this population (Chelouah and Siarry 2000). Also, large population sizes would be required to reduce the risk for solutions to get trapped in the local optima on the noisy and fluctuated error function surface. However, the drawback of large population sizes is the

36 Inverse analysis computation time cost. A small population would be more likely to converge quickly, which might be to a local optimum, and the quality of convergence is left to chance and in some cases could be equivalent to a random search. In the case of large population sizes, the population initialization is expensive in terms of computation time which in some cases could be equivalent to a random search.

A proper population size must be defined before running any genetic algorithm computations. An equation for choosing population size, from statistical theory proposed by Goldberg et.al (1991) and analyzed by him and Carroll (1996) has been studied in this research

= = (40) 푘 푘 푝표푝 푙 Where X is the number 푛of possibilities푚 푋 for�푘 each푋 chromosome, e.g. for binary X=2, is the length of chromosome and k is the size of the schema, details about the schema can be found in Goldberg (1989). 푙 To estimate a proper population size, it was considered that each parameter string would represent one schema. So, the length of the schema was assumed to be equal to the average parameter length for this purpose.

Strategy of running genetic algorithm

The search strategy is the combination of population size, offspring arrangement and termination criteria. The process of producing the initial population, following by breeding a new population continues until the termination conditions are met. In the process of optimization there is always a possibility that the next generation will produce a better solution which is the most important goal of optimization. On the other hand, the search could run forever and not be able to produce better solutions. Another situation could be that the solution’s set contains several similar fitness values. In this case engineering judgment is needed in order to select among them, since some combinations are more feasible and robust than others.

37 Identification of soil parameters in an embankment dam by mathematical optimization

38 Results

4 RESULTS

The purpose of this research is to introduce a new method in dam engineering for identifying the soil material properties within the dam structure (various zones) with the help of available measurement data of the instrumentations. For approaching this purpose, an inverse analysis method was introduced. In this study, the commercial finite element program Plaxis (2011) was used to simulate the earth and rockfill dam displacements. The optimizations were carried out by coupling Plaxis with a search algorithm program, based on the genetic algorithm, and minimizing an objective function of the least-square type.

In order to validate the inverse analysis technique and to examine the efficiency and robustness of the genetic algorithm, the research was initially focused on a synthetic case study. The synthetic case of model parameters, a set known in advance, is a good test for the mathematical basis used in the optimization, i.e. the objective function and the search algorithm. The Mohr- Coulomb elastic perfectly plastic model was chosen for all dam zones, as an initial choice for this research, chiefly because of its simplicity. A sensitivity analysis was utilized to characterize the influence of each parameter on displacements against the inclinometer data. From this analysis, the shear modulus of the core and the rockfill were found to be the most sensitive parameters for the horizontal deformation behaviour in the downstream zones of the dam in this study. A very good agreement was obtained in the optimization with the synthetic case. It was concluded from the synthetic case study, that the technique of inverse analysis could be effectively used in this kind of earth and rockfill dam application, presented in paper A. Table (2) summarizes the Mohr-Coulomb model parameters which are applied to this

39 Identification of soil parameters in an embankment dam by mathematical optimization

model. For the real dam case study, the simulated magnitudes of the horizontal displacements did not agree very well with the inclinometer measurements, see figure (24).

The results indicate that approaching a good agreement cannot be expected in an inverse analysis if the constitutive models chosen are not able to simulate the real soil behaviour sufficiently accurately. Moreover, a proper study of the genetic algorithm population size was needed since it has a large impact on the inverse analysis results.

According to the conclusions of the initial study, presented in paper A, it was clear that both models, the mathematical and the constitutive model, needed to be improved. For improving the mathematical model, population sizes of the genetic algorithm, for the synthetic case, were studied, see 3.9.2. The population size was chosen in such a way that each inverse computation almost converges to the same set of solutions. So, users feel confident about getting approximately the same results each time, running several inverse computations. Meanwhile, the importance of obtaining trustworthy solutions from an inverse analysis was judged against the need to save computation time. Therefore, a study of the population size as one of the key factors of the genetic algorithm efficiency was considered. For this purpose, fifteen inverse analyses were performed for the population sizes of 150, 120, 90, 75, 45 and 15. The results showed that by considering the importance of saving computation time and obtaining trustworthy solutions, the population size of 120 was the best population size for the synthetic case, presented in paper B.

Meanwhile, the Mohr-Coulomb model as a constitutive model had been improved by considering incremental shear modulus instead of constant shear modulus which was presented in paper A. By considering the incremental shear modulus as an optimization parameter the real behaviour of the soil can be captured better compared to a constant value of the shear modulus, since the stiffness of soil is assumed to increase linearly with depth with regard to the weight of soil.

Moreover, the topology of the error function was also studied for the synthetic case in order to establish a graphical shape of the error function, since only two parameters were considered to be optimized. Thereby, the smoothness properties of the surface can be examined as well as the location of the global optimum and the distribution of local optima in the search domain for the synthetic case. In this approach, presented in paper B, the surface was found to be non-smooth, with a lot of noises and fluctuations, and a large number of

40 Results local optima were spread over the entire search domain, see figure (21). It was concluded that the genetic algorithm is a proper choice as a search method for this rockfill dam application, since the genetic algorithm is not restricted by smoothness properties of the error function surface and solutions are not easily trapped into local optima, (presented in paper B).

The results of this approach with incremental shear moduli showed a decent agreement in optimization against the real inclinometer measurements, presented in paper B. Table (3) summarizes the Mohr-Coulomb model parameters, considering the incremental stiffness assumption, which is applied to this test. The simulated magnitudes of horizontal displacements agreed to some extent with the inclinometer measurements, see figure (24).

From the results of papers A and B, it was concluded that an advanced constitutive model would probably capture the real soil behaviour even better. Therefore, the Hardening soil constitutive model was chosen for the purpose of improving the inverse analysis. The Hardening soil model is an elasto-plastic cone-cap model, see chapter 3.8.1. Since, the mathematical model was significantly improved compared to previous studies, the inverse analysis only focused on the real dam case. Two different case studies (A and B), see chapter 3.5, based on different reservoir water levels and number of berms constructed, were considered. Case B is the same case which was monitored in previous studies. From another sensitivity analysis, the reference secant ref stiffness of the core material E50C and the reference secant stiffness of the ref rockfill material E50R were found to be the most sensitive parameters of the horizontal deformation behaviour in the downstream zones of the dam, for this constitutive model. Initially, in order to have a broad perspective on the location of the probable set within the domain, large limits were chosen for the ref ref optimization variables, E50C and E50R. Ten inverse analyses were utilized according to the range of optimization values. From the initial results it was concluded that the best set of solutions were located in a small range of ref ref E50R values, along the whole search range for E50C, values for both case A and case B. This appeared in a distinct valley in the surface of the error function as well. From these initial optimizations, a reduced search area was defined for case A and case B. After defining potential solution criteria from the first analyses, the second inverse analyses were applied to each case in the reduced domain. The set of solutions of the second ten inverse analyses were monitored in the domain of optimization parameters versus their relevant error function ref values. The results showed that the E50R value had changed during the time

41 Identification of soil parameters in an embankment dam by mathematical optimization

ref (from case A to case B) while most probably the E50C value had not changed that much, presented in paper C.

The topology of the objective function for case A was found not to be as fluctuating and noisy as the topology obtained with the Mohr-Coulomb model. Again from this case study it was concluded that the genetic algorithm is a proper search algorithm for this kind of application since it is not restricted to smoothness properties of the error function. On the other hand, the topology of case B was found to be smooth without noticeable fluctuations. The error function shape in this study based on the Hardening soil model, had significantly changed compared to the shape obtained in the previous study which was very fluctuating and noisy, based on the Mohr-Coulomb model.

A very good agreement from the best set of solutions was obtained with the inclinometer measurements for both case A and case B. Table (4) summarizes the Hardening soil model parameters which are utilized for this case. The Hardening soil model captured the pattern of horizontal displacements quite well. On the other hand, finding a proper range of values in the domain of ref ref [E50C, E50R] had a significant influence on the improvement of the results, see figures (24) and (25).

According to figure (24), since the Hardening soil model is a stress-dependent model, it is able to capture the horizontal displacements dam crest perfectly, while the Mohr-Coulomb model is not stress dependent and was not able to capture the horizontal displacements on the crest. On the other hand, in the incremental stiffness version of the Mohr-Coulomb model the stiffness values increased with depth, and this model showed quite similar displacement behaviour in the bottom of the dam compared to the Hardening soil model.

42 Results

Figure 21: Error function and level contours of error function for Mohr- Coulomb model, presented in paper B

43 Identification of soil parameters in an embankment dam by mathematical optimization

Figure 22: Error function topology and level contours for case A of the Hardening soil model

44 Results

Figure 23: Error function topology and level contours for case B of the Hardening soil model

45 Identification of soil parameters in an embankment dam by mathematical optimization

Inclinometer values MC model values MC increm-stiffness model values Hardening soil model values 50

45

40

35

30

25

20

The height ofthe dam [m] 15

10

5 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 Horizontal displacement [m] Figure 24: Horizontal displacement versus the height of the dam for different constitutive models

Case A- inclinometer FE model results of case A Case B- inclinometer FE model results of case B 45

40

35

30

25

20

15

The height of the dam [m] 10

5 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 Horizontal displacement [m] Figure 25: Measured and computed horizontal displacements at different heights of the dam for cases A and B

46 Results

Table 2: Material properties of the rockfill dam (Mohr-Coulomb model-presented in paper A)

γu γs G'ref ´ ´ ϕ´ kx/ky Dam zones kN/m3 MPa 휈- kPa푐 (°) m/s

Core 21 23 31 0.35 20 38 3.0·10-7

Fine filter 21 23 64 0.33 0 32 9.0·10-5

Coarse filter 21 23 64 0.33 0 34 5.0·10-4

Rockfill 19 21 15 0.33 7 30 1.0·10-2

Berms 21 23 37 0.33 7 36 5.0·10-2

Foundation 21 23 3700 0.30 0 45 1.0·10-8

Table 3: Material properties of the rockfill dam (Mohr-Coulomb model, presented in paper B)

γu γs G'ref G'inc ν' c' ϕ' kx/ky Dam zones kN/m3 MPa kPa - kPa (°) m/s

Core 21 23 18 524 0.35 20 38 3.0·10-7 Fine filter 21 23 40 100 0.33 0 32 9.0·10-5 Coarse 21 23 40 100 0.33 0 34 5.0·10-4 filter Rockfill 19 21 10 60 0.33 7 30 1.0·10-2

Berms 21 23 13 0 0.33 7 30 5.0·10-2

Foundation 21 23 3700 0 0.30 0 45 1.0·10-8

Note: γ u is the unit weight above the phreatic level,γ s is the unit weight below the phreatic level, G'ref is the shear modulus, G'inc is the shear modulus increment, ν' is the effective Poisson’s ratio, c' is the effective cohesion, φ' is the effective friction angle and kx and ky are the hydraulic conductivity in horizontal and vertical direction, respectively.

47 Identification of soil parameters in an embankment dam by mathematical optimization

Table 4: Material properties of the rockfill dam (Hardening soil model, presented in paper C)

ref ref ref γu γs E50 =Eoed Eur m c' ϕ' kx/ky Dam zones kN/m3 MPa MPa - kPa (°) m/s

Core 21 23 70-100 210-300 1 20 38 3.0·10-7

Fine filter 21 23 50 150 0.5 0 32 9.0·10-5

Coarse filter 21 23 50 150 0.5 0 34 5.0·10-4

Rockfill 19 21 10-17 30-51 0.5 7 30 1.0·10-2

Berms 21 23 10 30 0.5 7 30 5.0·10-2

Foundation 21 23 3000 9000 0.5 0 45 1.0·10-8

Note: γ u is the unit weight above the phreatic level,γ s is the unit weight ref below the phreatic level, E50 is the Reference secant stiffness for primary ref loading in a drained triaxial test, Eoed is the Reference tangent stiffness ref for primary oedometer loading, Eur is the Reference stiffness for unloading\reloading in a drained triaxial test, m is the Power for stress- level dependency of stiffness, c' is the effective cohesion, ϕ' is the

effective friction angle and kx and ky are the hydraulic conductivity in horizontal and vertical direction, respectively.

Finally the overall outcome of the identification of soil material properties of the earth and rockfill dam by inverse analysis computations can be listed as follows:

• The inverse analysis technique is a proper non-destructive method for soil parameter identification of earth and rockfill dams.

• The inverse model consists of two models; an optimization model (search algorithm and error function) and a numerical model. Therefore, a good inverse analysis highly depends on the choice of a proper mathematical and numerical model.

• The proper choice of a constitutive relation, in order to capture the real soil behaviour is one of the key factors of performing an inverse

48 Results

analysis. Results with the Hardening soil model show that this model could be a proper choice for doing inverse soil parameter identification in earth and rockfill dam applications.

• The genetic algorithm is a robust and efficient search algorithm for finding a set of solutions close to the global optimum, without being trapped into local optima. The choice of a proper population size is one of the key factors in the genetic algorithm. So, an initial study needs to be made in order to find the best population size regarding the problem.

• Considering a proper range of variation of optimization parameter values has an impact influence on finding the best set of solutions as close as possible to the global optimum solution.

49 Identification of soil parameters in an embankment dam by mathematical optimization

50 Future work

5 FUTURE WORK

Future work can be focused on testing the mathematical basis against field measurements of different kinds, like: pore pressures, total stresses, seepage etc. In this research, only horizontal displacements from inclinometer measurements were used. Identification of constitutive parameter values might be improved if more measurements as well as measurements of different kinds are included in the error function.

Moreover, evaluating the mathematical basis with a larger number of optimization variables and their correlations can be studied in the future.

In order to evaluate the introduced inverse analysis method; more case studies with more measurements data can be tested.

A more in detail study of the influence on optimization of the genetic algorithm operations such as mutation and cross over can be considered for future work.

51 Identification of soil parameters in an embankment dam by mathematical optimization

52 References

REFERENCES

1- Bowles, J.E. (1988). "Foundation Analysis and Design". MacGraw-Hill. Berkshire. England. 2- Calvello, M. and Finno, R., (2002). "Calibration of soil models by inverse analysis". Numerical Models in Geomechanics NUMOG VIII, pp. 107-116. 3- Calvello, M. and Finno, R.J., (2004). "Selecting Parameters to Optimize in Model Calibration by Inverse Analysis". Computers and Geotechnics, vol. 31, no. 5, pp. 410-424. 4- Carroll, D.L., (1996). "Genetic algorithms and optimizing chemical oxygen- iodine lasers". Developments in Theoretical and Applied Mechanics, vol. 18, no. 3, pp. 411-424. 5- Chelouah, R. and Siarry, P., (2000). "A continuous genetic algorithm designed for the global optimization of multimodal functions". Journal of Heuristics, vol. 6, no. 2, pp. 191-213. 6- Chen, W. F. and Mizuno, E., (1990). "Nonlinear analysis in soil mechanics: theory and implementation". 53rd ed. Amsterdam: Elsevier. 7- Chen, W., (2007). "Plasticity in reinforced concrete". J. Ross Publishing. 8- Clough, R.W. and Woodward, R.J., (1967). "Analysis of Embankment Stresses and Deformations". Journal of Soil Mechanics & Foundations Div. Vol 93, pp. 529-549. 9- Coulomb, C.A., (1776). "Essai sur une application des règles de maximis & minimis à quelques problèmes de statique, relatifs à l'architecture". De l'Imprimerie Royale. 10- Drucker, D.C., (1950). "Stress-Strain Relations in the Plastic Range". A Survey of Theory and Experiment.

53 Identification of soil parameters in an embankment dam by mathematical optimization

11- Drucker, D.C. and Prager, W., (1952). "Soil mechanics and plastic analysis or limit design". Quarterly of Applied Mathematics, vol. 10. 12- Drucker, D.C., Gibson, R.E. and Henkel, D.J., (1957). "Soil mechanics and work hardening theories of plasticity". Transactions of the American Society of Civil Engineers, vol. 122, pp 338-346. 13- Duncan, J.M. and Chang, C.Y., (1970). "Nonlinear analysis of stress and strain in soils". Journal of the Soil Mechanics and Foundations Division, vol. 96, no. 5, pp. 1629-1653. 14- Duncan, J.M., (1981). "Hyperbolic Stress-Strain Relationships". Application of Plasticity and Generalized Stress-Strain in Geotechnical Engineering. 15- Duncan, J.M., (1996). "State of the Art: Limit Equilibrium and Finite-Element Analysis of Slopes". Journal of Geotechnical Engineering, vol. 122, no. 7, pp. 577-596.

16- Finno, R.J. and Calvello, M., (2005). "Supported Excavations: Observational Method and Inverse Modeling". Journal of Geotechnical and Geoenvironmental Engineering, vol. 131, no. 7, pp. 826-836. 17- Gallagher, K., Sambridge, M.S. and Drijkoningen, G., (1994). "Genetic Algorithms-an Evolution from Monte-Carlo Methods for Strongly Non-Linear Geophysical Optimization Problems". Geophysical Research Letters, vol. 18, no. 12, pp. 2177-2180. 18- Gioda, G. and Maier, G., (1980). "Direct Search Solution of an Inverse Problem in Elastoplasticity: Identification of Cohesion, Friction Angle and in Situ Stress by Pressure Tunnel Tests". International Journal for Numerical Methods in Engineering, vol. 15, no. 12, pp. 1823-1848. 19- Gioda, G. and Sakurai, S., (1987). "Back Analysis Procedures for the Interpretation of Field Measurements in Geomechanics". International Journal for Numerical and Analytical Methods in Geomechanics, vol. 11, no. 6, pp. 555-583. 20- Goldberg, D.E., (1989). "Genetic algorithms in search, optimization, and machine learning". Machine learning 3.2 (1988): 95-99 21- Goldberg, D.E., Deb, K. and Clark, J.H., (1991). "Genetic algorithms, noise, and the sizing of populations". Complex Systems, vol. 6, pp. 333-362. 22- Hill, R., (1998). "The mathematical theory of plasticity". Oxford university press. 23- Hunter, G. and Fell, R., (2003). "The deformation behaviour of embankment dams". Uniciv report No. R-416. The University of New Wales, School of Civil and Environmental Engineering.

54 References

24- Kennett, B. L., and Sambridge, M. S. (1994). "Hypocentre location: genetic algorithms incorporating problem specific information". Geophysical Journal International, 118(3), 693-706. 25- Koiter, W.T., (1960). "General theorems for elastic-plastic solids". North- Holland Amsterdam. 26- Kovacevic, N., (1994). "Numerical Analyses of Rockfill Dams, Cut Slopes and Road Embankments". PhD Thesis, London University (Imperial College of Science and Medicine), Faculty of Engineering.

27- Ledesma, A., Gens, A. and Alonso, E., (1996). "Estimation of Parameters in Geotechnical back analysis—I. Maximum Likelihood Approach". Computers and Geotechnics, vol. 18, no. 1, pp. 1-27. 28- Lecampion, B., Constantinescu, A. and Nguyen Minh, D., (2002). "Parameter Identification for Lined Tunnels in a Viscoplastic Medium". International Journal for Numerical and Analytical Methods in Geomechanics, vol. 26, no. 12, pp. 1191-1211. 29- Lecampion, B. and Constantinescu, A., (2005). "Sensitivity Analysis for Parameter Identification in quasi‐static Poroelasticity". International Journal for Numerical and Analytical Methods in Geomechanics, vol. 29, no. 2, pp. 163-185. 30- Levasseur, S., Malecot, Y., Boulon, M. and Flavigny, E., (2008). "Soil Parameter Identification using a Genetic Algorithm". International Journal for Numerical and Analytical Methods in Geomechanics, vol. 32, no. 2, pp. 189- 213. 31- Levasseur, S., Malecot, Y., Boulon, M. and Flavigny, E., (2009). "Statistical Inverse Analysis Based on Genetic Algorithm and Principal Component Analysis: Method and Developments using Synthetic Data". International Journal for Numerical and Analytical Methods in Geomechanics, vol. 33, no. 12, pp. 1485-1511. 32- Levasseur, S., Malecot, Y., Boulon, M. and Flavigny, E., (2010). "Statistical Inverse Analysis Based on Genetic Algorithm and Principal Component Analysis: Applications to Excavation Problems and Pressuremeter Tests". International Journal for Numerical and Analytical Methods in Geomechanics, vol. 34, no. 5, pp. 471-491. 33- Marsal, R.J. and De, A.L.R., (1900). "Performance of El Infiernillo Dam, 1963-1966". Journal of Soil Mechanics & Foundations Div, vol. 92, no. SM5, Proc Paper 490. 34- Marsal, R.J., (1959). "Earth Dams in Mexico" Anonymous Proceedings, 1st Panamerican Conference on Soil Mechanics and Foundation Engineering, Mexico, DF.

55 Identification of soil parameters in an embankment dam by mathematical optimization

35- Marsal, R. and Fuentes de la Rosa, A., (1976). "Mechanical properties of rockfill soil mixtures". Anonymous Transactions of the 12th International Congress on Large Dams, Mexico City.

36- Melan, E., (1938). "Zur plastizität des räumlichen kontinuums". Archive of Applied Mechanics, vol. 9, no. 2, pp. 116-126. 37- Mises, R.v., (1928). "Mechanik der plastischen Formänderung von Kristallen". Zamm Journal of Applied Mathematics and Mechanics/Zeitschrift Für Angewandte Mathematik Und Mechanik, vol. 8, no. 3, pp. 161-185. 38- Moreira, N., Miranda, T., Pinheiro, M., Fernandes, P., Dias, Daniel., Costa, L., Sena-Cruz, José. (2013). "Back analysis of geomechanical parameters in underground works using an Evolution Strategy algorithm". Tunnelling and Underground Space Technology. vol. 33, pp.143-158, Elsevier 39- Muir Wood, D., (2004). "Geotechnical Modelling". Taylor & Francis. Spon Press. Oxfordshire.

40- Naghdi, P.M., (1960). "Stress-Strain Relations in Plasticity and Thermoplasticity". No. TR-9. California University Berkeley institute of engineering research. 41- Nelder, J.A. and Mead, R., (1965). "A simplex method for function minimization". The Computer Journal, vol. 7, no. 4, pp. 308. 42- Nobari, E. and Duncan, J.M., (1972). "Effect of Reservoir Filling on Stresses and Movement in Earth and Rockfill Dams". A Report of an Investigation. No. TE-72-1. California University Berkeley coll of engineering.

43- Papon, A., Riou, Y., Dano, C. and Hicher, P. (2012). "Single- and multi- objective genetic algorithm optimization for identifying soil parameters. International Journal for Numerical and Analytical Methods in Geomechanics, vol. 36, no. 5, pp. 597-618. 44- Plaxis. (2011). Finite Element Code for Soil and Rock Analysis. version 2011- 2.http://plaxis.nl 45- Potts, D.M. and Zdravković, L., (2001). "Finite Element Analysis in Geotechnical Engineering: Application". Vol.2. Thomas Telford Services Limited, London.

46- Prager, W., (1955). "The theory of plasticity: a survey of recent achievements". Proceedings of the Institution of Mechanical Engineers, vol. 169, no. 1, pp. 41-57. 47- Rankine, W.M., (1857). "On the stability of loose earth". Philosophical Transactions of the Royal Society of London, pp. 9-27.

56 References

48- Rechea, C., Levassur, S. and Finno, R., (2008). "Inverse Analysis Techniques for Parameter Identification in Simulation of Excavation Support Systems". Computers and Geotechnics, vol. 35, no. 3, pp. 331-345. 49- Resendiz, D. and. Romo, M.P (1972). "Analysis of embankment deformations". Anonymous Performance of Earth and Earth-Supported Structures. ASCE. Vol.1. pp. 817-836 50- Roscoe, K. and Burland, J., (1968). "On the generalized stress-strain behaviour of wet clay". 51- Rowe, P.W., (1962). "The stress-dilatancy relation for static equilibrium of an assembly of particles in contact". Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences, vol. 269, no. 1339, pp. 500-527. 52- Schanz, T., P. Vermeer and P. Bonnier., (1999). "The hardening soil model: formulation and verification" Anonymous Beyond 2000 in computational geotechnics: 10 years of PLAXIS International; proceedings of the International Symposium beyond 2000 in Computational Geotechnics, Amsterdam, The Netherlands, 18-20 March 1999, 1999. 53- Schofield, A.N. and Wroth, P., (1968). "Critical state soil mechanics". 54- Sharma, H. D., G. C. Nayak, and J. B. Maheshwari., (1979). "Nonlinear analysis of a high rockfill dam with earth core." 13th International Congress on Large Dams, ICOLD, New Delhi, India. Vol. 1. 55- Sherard, J.L., (1953). "Influence of soil properties and construction methods on the performance of homogeneous earth dams". US Department of the Interior, Bureau of Reclamation, Design and Construction Division. 56- Shield, R., (1955). "On Coulomb's law of failure in soils". Journal of the Mechanics and Physics of Solids, vol. 4, no. 1, pp. 10-16. 57- Stoffa, P.L. and Sen, M.K., (1991). "Nonlinear multiparameter optimization using genetic algorithms; inversion of plane-wave seismograms". Geophysics, vol. 56, no. 11, pp. 1794. 58- Stroyizdat Moscow, (1993). "Rock materials for rockfill dams; review and recommendations". Bulletin - Commission Internationale Des Grands Barrages = Bulletin - International Commission on Large Dams, vol. 92 GeoRef. ISSN 0534-8293, 0534-8293. 59- Sundaram, A., and Venkat Venkatasubramanian., (1998). "Parametric sensitivity and search-space characterization studies of genetic algorithms for computer-aided polymer design." Journal of chemical information and computer sciences 38.6:1177-1191.

57 Identification of soil parameters in an embankment dam by mathematical optimization

60- Swoboda, G., Ichikawa, Y., Dong, Q. and Zaki, M., (1999). "Back Analysis of Large Geotechnical Models". International Journal for Numerical and Analytical Methods in Geomechanics, vol. 23, no. 13, pp. 1455-1472. 61- Tarantola, A., (2005). "Inverse Problem Theory". Elsevier. Amsterdam 62- Tatsuoka, F. and Ishihara, K., (1974). "Yielding of sand in triaxial compression. Soils and Foundations". vol. 14, no. 2, pp. 63-76. 63- Terzaghi, K., (1958). "Design and performance of the Sasumua Dam". Harvard University, Division of Engineering and Applied Physics. 64- Vermeer, P., (1978). "A double hardening model for sand". Geotechnique, vol. 28, no. 4, pp. 413-433. 65- Wilde, P., (1977). "Two-invariant dependent model of granular media". Arch.of Mechanics, vol. 29, pp. 799-809. 66- Wood, D.M., (1990). "Soil behaviour and critical state soil mechanics". Cambridge university press. 67- Yu, Y., Zhang, B. and Yuan, H., (2007). "An Intelligent Displacement Back- Analysis Method for Earth-Rockfill Dams". Computers and Geotechnics, vol. 34, no. 6, pp. 423-434. 68- Yu, H., (2006). "Plasticity and geotechnics". Vol. 13. Springer. 69- Yu, H., (1998). CASM: "A unified state parameter model for clay and sand". International Journal for Numerical and Analytical Methods in Geomechanics, vol. 22, no. 8, pp. 621-653. 70- Yu, H., Khong, C., Wang, J. and Zhang, G., (2005). "Experimental evaluation and extension of a simple critical state model for sand". Granular Matter, vol. 7, no. 4, pp. 213-225. 71- Yu, Y., Zhang, B. and Yuan, H., (2007). "An Intelligent Displacement Back- Analysis Method for Earth-Rockfill Dams". Computers and Geotechnics, vol. 34, no. 6, pp. 423-434. 72- Zentar, R., Hicher, P. and Moulin, G., (2001). "Identification of Soil Parameters by Inverse Analysis". Computers and Geotechnics, vol. 28, no. 2, pp. 129-144.

58

Paper A

Inverse soil parameter identification of an earth and rockfill dam by genetic algorithm optimization

Vahdati, P., Levasseur, S., Mattsson, H. and Knutsson, S

In proceedings of the 9th ICOLD European club symposium, Sharing Experience for Safe and Sustainable Water Storage, 10-12 April 2013 Venice, Italy, Topic B

Inverse soil parameter identification of an earth and rockfill dam by genetic algorithm optimization

Pooya Vahdati1, Séverine Levasseur2, Hans Mattsson1, Sven Knutsson1

1Luleå University of Technology, 97187 Luleå-Sweden, 2 Université de Liège, 4000 Liège-Belgium E-mail: [email protected]

Summary consist of various soil types for which the stress/strain response could vary considerably. For each dam zone, suitable values The paper presents a study of constitutive parameter value have to be assigned to the parameters included in the identification by inverse analysis on an earth and rockfill dam constitutive model chosen. In general, laboratory tests and/or application. The objective is to examine if the technique of field tests of the soil are needed as a basis for this parameter inverse analysis can be effectively used for this type of case. In evaluation. However, many dams are old and limited the inverse analysis procedure discussed here, values of information might be available regarding the soil materials constitutive parameters are determined based on data recorded being used in the dam structures. In dams, it is normally very from installed instrumentations in the dam. The quantities that difficult to take up soil samples for testing, especially from the are monitored in the dam can be numerically predicted by a central impervious part, since this might affect the dam finite element simulation. To perform a finite element performance and the safety of the dam. For dams it would be simulation, constitutive models have to be chosen and values advantageous if constitutive parameter values could be have to be assigned to the parameters included. In inverse determined with some non-destructive method. analysis, constitutive parameter values are chosen in such a Inverse analysis provides a possibility to determine the way that the error between data obtained by measurements in constitutive behavior of different materials within the dam the dam and numerical simulation is minimized. This is structure under the condition that the dams have been equipped accomplished by optimization. The genetic algorithm was with various instrumentations, for monitoring dam utilized as the optimization strategy, to search for the minimum performance, which record data such as pore pressures, error, in this study. Optimizations have been performed both deformations, total stresses and seepage etc. In the proposed against a synthetic dam case and a real dam case. It was inverse analysis, also named back analysis, a finite element concluded, that the inverse analysis technique studied, could be (FE) model is calibrated automatically by changing the values effectively used for this kind of earth and rockfill dam of the input parameters of the selected constitutive model in application. However, the technique was time consuming. different dam zones until the discrepancy between measured Inverse analysis has the potential to become a valuable tool for results by dam instrumentations and the corresponding dam engineers assessing dam performance and dam safety if it computed results is minimized. The finite element model becomes readily available in commercial finite element calibration is conducted by applying an optimization method to software. search for the best possible solution. The inverse analysis technique was introduced to the geotechnical field by Gioda [5] with the purpose of identification of elastic material properties Introduction of in situ rock masses. The inverse analysis method has since then been applied in many different types of geotechnical The finite element method (FEM) has been widely used to applications [6]-[15]. analyze earth and rockfill dams. References [1] and [2] are both The objective of this paper is to examine if the technique of examples of good state of the art papers on finite element inverse analysis can be effectively used for identification of analysis of deformation behavior in earth and rockfill dams. constitutive parameter values in an earth and rockfill dam The books by Potts and Zdravkovic' [3] and Muir Wood [4] application. An inverse analysis method for dam applications contain many valuable general advices about how numerical was proposed in reference [16], neural networks were utilized modeling of e.g. dams should be performed. In a finite element which is a different technique compared to the finite element analysis a proper constitutive model has to be chosen for each approach used in this paper. If the values of the constitutive part of the dam in order to simulate the relation between parameters can be identified by inverse analysis, reliable finite stresses and strains. The zones of an earth and rockfill dam element simulations of the dam structure can be performed. The have different functions. Because of that the zones normally results from such simulations can be interpreted and put in

1 relation to dam monitoring data observed to obtain a better be tested against the challenge of a gradually increased number overall understanding of the mechanical behavior of the dam. of optimization variables. Dam performance as well as dam safety can then be better assessed. In this study, geometry and zoning as well as The earth and rockfill dam inclinometer data from an existing earth and rockfill dam has been used, see figure (1). The earth and rockfill dam used for the case study is 45m high, In the technique of inverse analysis an optimization method with a dam crest of 6.5m, see figure (1). The upstream slope is is needed. The mathematical procedure of optimization inclining 1V:1.71H and the downstream slope is inclining basically consists of two parts, the formulation of an objective 1V:1.85H. The impervious core is surrounded by fine and function measuring the difference between monitored data and coarse filters and supporting fill of blasted rock. After some the associated data simulated by finite element analysis, and the time, the dam was strengthened in two stages with rockfill selection of an optimization strategy enabling the search for the berms on the downstream part. Horizontal displacements are minimum of this objective function. Horizontal deformations measured automatically by inclinometers installed along the are the main concern for the earth and rockfill dam in this downstream filter, from the crest into the rock foundation. study. Therefore, the objective function has been formulated to Since the inclinometer data from the ten first meters from the give the difference between horizontal deformations obtained bottom of the core seems to be not trustworthy, only data above by inclinometer measurements and finite element simulations, this level are considered. The level of water in the reservoir respectively. As the search strategy, the genetic algorithm (GA) varied from +430m to +440m during the one year of interest in has been chosen in this study. The genetic algorithm is a this study. stochastic search method which is well suited in finding near optimal solutions to complex problems, but without any guarantee of finding the exact global optimum [17]-[19]. This search strategy has been chosen due to its robustness, efficiency and particularly for providing a set of solutions close to the optimum solution instead of a unique answer; which is more practical from a geotechnical perspective. This method has been utilized in several geotechnical applications, see e.g. reference [13], [20] - [22]. However, to the authors’ knowledge, Figure 1 : Cross section of the dam the genetic algorithm has still not been used in earth and rockfill dam applications. The practical outcome of an inverse analysis clearly depends Finite element model on the ability of the constitutive models chosen, to mimic the real soil behavior in the different dam zones. To evaluate the The Plaxis finite element program [23] was utilized for the robustness and efficiency of the genetic search algorithm for numerical analysis in this study. The finite element model of the dam application, the uncertainty of the prediction ability of the cross section of the dam is shown in figure (2). Since the the constitutive model chosen has been removed from the dam is a long structure, plain strain condition was assumed. analysis by optimizing against deformations obtained by finite Some smaller simplifications compared to the real dam cross element simulations with an in advance known set of model section in figure (1) were done in the finite element model in parameters. A synthetic case has been created. The actual figure (2). These simplifications are considered to be of no optimization problem is then invented by choosing other sets of practical importance in this study. For an accurate analysis, a model parameters as starting-points for the optimization. sufficient large part of the cross sectional geometry was Thereby, the optimum is known in advance and it is no doubt considered in the finite element model. The option standard that it can be reproduced again. fixities in Plaxis was selected for generation of boundary In this study, the genetic algorithm has been validated first conditions to the geometry model, for the outermost vertical against a synthetic case. After that, model parameter calibration left and right foundation boundary horizontal displacements has been performed against real inclinometer measurements on were prescribed to zero and for the lowest horizontal boundary the dam. The authors’ are in the initial stages of the project in in the foundation both horizontal and vertical displacements which this research results have been produced. The were prescribed to zero. The element type was chosen as 15- optimizations presented in this paper are restricted to two node triangular elements. The coarseness of the mesh was optimization variables. In the future, the genetic algorithm will selected so that a sufficient accuracy of the computations was obtained for a minimum of computation time spent. Several

2 different meshes were tested and the one illustrated in figure (2) was chosen. The dam embankment in the model in figure (2) was constructed stepwise in five stages in order to generate a proper initial effective stress field [24], [25]. The excess pore pressures in the core were dissipated by consolidation after each stage. The Mohr-Coulomb elastic perfectly plastic model was chosen for all dam zones, as an initial choice for this research, foremost because of its simplicity. Table (1) summarizes the Mohr-Coulomb model parameters which are applied for this case study. The parameter values of ȖuȖsij', kx and ky were provided by the dam owner. Appropriate values of Figure 2: Finite element model of cross section of the dam: Ȟ and c' were chosen based on advices in reference [26]. The 2D model in plain strain dilatancy angle were calculated from the expression ȥ' ij'-30° in reference [23]. A high value of the shear modulus was ܩ௥௘௙ TABLE 1 : MATERIAL PROPERTIES OF THE ROCKFILL DAM selected for the foundation to make it rigid. The values of the shear modulus, ܩ for all the other materials in the dam were ௥௘௙ Dam Ȗu Ȗs Gref ߥ´ ܿ´ ߮´ kx/ky identified by calibration against inclinometer results. The dam zones core was modeled as undrained case (B) in the Plaxis software. kN/m3 MPa - kPa (°) m/s The undrained shear strength needed in case (B) was assumed -7 Core 21 23 31 0.35 20 38 3.0·10 to increase linearly with depth and was calculated with the effective parameters ߮ᇱ and ܿᇱ for representative effective stress F.filter 21 23 64 0.33 0 32 9.0·10-5 states at the top and the bottom of the core, before C.filter 21 23 64 0.33 0 34 5.0·10-4 impoundment. All the other dam zones were modeled with the option drained behavior, i.e. no excess pore pressures are Rockfill 19 21 15 0.33 7 30 1.0·10-2 generated. Berms 21 23 37 0.33 7 36 5.0·10-2 In the finite element analyses executed in this paper, -8 simulated horizontal deformations received in exactly the same Foundation 21 23 3700 0.30 0 45 1.0·10 positions in the geometry as the measurements carried out with Note: ߛ௨ is the unit weight above phreatic level, ߛ௦ is the unit ᇱ the single inclinometer in figure (1), have been compared to the weight below phreatic level, ܩ௥௘௙ is the shear modulus, ߥ is the corresponding inclinometer measurements in the expression for effective Poisson’s ratio, ܿᇱ is the effective cohesion, ߮ᇱ is the the objective function. The inclinometer measurements in the effective friction angle, ȥ is the dilatancy angle and kx and ky real dam began when the reservoir was filled for the first time. are the hydraulic conductivity in horizontal and vertical Thus, in the finite element analyses, all displacements were direction, respectively. reset to zero before impoundment. The deformation comparison in this study has been performed for the specific moment in time when the water level in the reservoir was at the level Optimization method +440m and the construction of the berms in figure (1) were just completed. To reach this particular state in the numerical The aim of optimization is to approach the optimum solution analysis, the reservoir was filled in four stages from +400m to among other solutions for a given problem in a matter of +440m and the water pressure was generated based on the efficiency. For optimization an objective function and a search phreatic level representative for a steady state water flow method are needed. In this paper, the objective function should condition. After that the berms were constructed. The excess provide a scalar measure of the deviations between horizontal pore pressures in the core were dissipated before construction deformations obtained by inclinometer measurements and of the berms, but not afterwards. numerical simulation, respectively. As the search method, for finding the minimum value of this objective function, the genetic algorithm (GA) is chosen. The variables to optimize consist of selected parameters included in constitutive models utilized for the finite element simulations. Notice that, the magnitude of the objective function is thereby influenced by the values of the optimization variables and the actual problem is to

3 find the set of values of the optimization variables that on the individual’s fitness value the population is sorted in minimizes the objective function; that is the optimum solution. ascending or descending order. With the aim of collecting the best individual, only Ni/3 of them are retained for the next Objective function population, which are called parents. Randomly selected pair In order to define the difference between the measured values parents produce a new generation (offspring) based on and FE computed values, it is essential to define a proper crossover or mating. The most common form of mating objective function. In this research, following Levasseur et al. includes two parents that produce two off springs. The position [20], a scalar error function Ferr based on the least-square of crossover points is randomly chosen and the value of each method was introduced binary string exchange with another one after this position point, see figure (3). In order to improve the efficiency of the ଵ ே ଶ ൗଶ algorithm a crossover point number is chosen equal to the ൫ܷ െܷ ൯ 1 ௘೔ ௡೔ (1) number of selected parameters, which was proposed by Pal et ܨ௘௥௥ = ൭ ൗܰ ෍ ଶ ൱ ߂ܷ al. [17]. This process is repeated until 2Ni/3 off springs ௜ୀଵ ௜ (children) are created. Now the new population, consisting of that determines the difference between experimental measured parents and children is completed. Mutations apply for preventing the GA from converging too fast before sweeping values Uei and numerically computed values Uni, from the N the entire search domain. Therefore some of the individuals are measuring points. The term ǻ8i gives the weight on the randomly selected to flip their character, see figure (4). discrepancy between Uei and Uni. The parameter ǻ8i relates to the numerical and experimental uncertainties at the measurement point “i” defined as

Parent #1 111001 100100 111001011011 offspring#1 ߂ܷ = ߝ + ߙܷ (2) ௜ ௘೔ Parent #2 001110 011011 001110100100 offspring#2 where the parameter ߝ is an absolute error of measurements and parameter ߙ is a dimensionless relative error of measurements. Figure 3: Reproduction of offspring (children) Search algorithm from the pair of parents The genetic algorithm that belongs to the subcategory of stochastic search methods was developed by John Holland [27] in 1970s at the University of Michigan; that method was inspired by Darwin’s theory of evolution. GA is a global optimization technique which is based on randomized operators Selected gene before mutation 001011010110 such as: selection, crossover and mutation. The search method Selected gene after mutation 001010010110 can approach to a set of solutions close to the optimum solution Figure 4: Mutation of an individual without being trapped into local optimums in the given search domain. The genetic algorithm is briefly described below. A more thorough description can e.g. be found in the book by Optimization results of the synthetic case Haupt [19]. The optimization program that is used in this research was developed in the Fortran language by Levasseur Before applying the inverse analysis on the real inclinometer et.al [13], [20] for geotechnical studies. measurements performed on the earth and rockfill dam in figure In this genetic algorithm, the area of search is defined based (1), the robustness and efficiency of the optimization method on the number of parameters Np to optimize. The minimization are validated against a synthetic case. The inclinometer has the problem is approached in the Np dimensional space limited by position as illustrated in figure (1). The results of the the upper and the lower bound constrains of each parameter. inclinometer measurements are in the synthetic case obtained The values of the optimization parameters are binary encoded from a finite element simulation of the dam with the Mohr- in a form of genes. The combination of the genes forms Coulomb model and the parameters in table (1). By this individuals Ni. The population (chromosomes) is made by strategy, it is an evident fact that the Mohr-Coulomb model is several individuals in the domain. The initial population is able to simulate the synthetic case and the actual optimization formed based on a random selection in the search space. Based

4 is started from a specified search domain for model parameters able to find relevant set of solutions close to the optimum including the initial set (optimal solution) defined in table (1). values. Before optimization, a sensitivity analysis of the model parameters in table (1) has been performed to characterize the Initial population influence of each parameter (not presented in this paper). From 55000 this it was found that the shear modulus Gref of the core and the shear modulus Gref of the rockfill material were two of the most 50000 sensitive parameters for the horizontal deformation behavior in [kPa] [kPa] the downstream zones of the dam. These two parameters were 45000 ref selected as optimization variables in this study. The search G 40000 domain was limited by lower and upper bounds for each Initial core 35000 population optimization variable as: 29MPa ” G ref ” 93MPa and rockfill Optimum 12MPa ” G ref ” 56MPa. The optimum solution in this 30000 core solution synthetic case is known as the values G ref =31MPa, rockfill 25000 G ref =15MPa. The optimization results of the synthetic case are shown in 20000 figure (5), (6) and (7). In figure (5), the initial population Shear modulus of rockfill 15000 generated by the genetic algorithm, is presented. The initial population represents the start of the search for the optimum, 10000 and consists of 90 points in the two dimensional search space 25000 40000 55000 70000 85000 100000 spanned by the optimization variables and their boundaries. Shear modulus of core Gref [kPa]

For each of the following populations, half of the number of points in the initial population is applied (45 points). The error Figure 5 : Initial population function defined in equation (1) is considered with the SDUDPHWHUVİ DQGĮ $IWHUQLQHLWHUDWLRQVRQHVHWRI solutions around the optimum solution is identified; see figure Solution 9th population (6) and table (2). From this set of solution, the user can choose 18000 a solution with a low value of the error function and a set of parameter values that “feels correct” from engineering 17000 perspective.

To go further in the analysis and to have an idea of genetic [kPa] 16000 algorithm capabilities, we can compare different numerical ref G results. In figure (7), horizontal displacements versus the height 15000 Solution 9th population of the dam are shown for the inclinometer position in the dam. 3 1 The inclinometer values from the synthetic case represent the 14000 Optimum optimum solution. On this figure, numerical curves obtained solution from several parameter sets are compared. First, we consider a 13000 ௖௢௥௘ rockfill curve obtained from ܩ௥௘௙ = 84MPa, and G ref =53MPa, in 12000 the initial population for which the error function is equal to 2 5E-2. Thereafter, we consider curves obtained from the of rockfill modulus Shear identified solution set; e.g. points 1 and 2 which give the 11000 smallest error function values and correspond to shear moduli ௖௢௥௘ rockfill ௖௢௥௘ 10000 ܩ௥௘௙ = 31MPa and G ref =14.5MPa for point 1 and ܩ௥௘௙ = 25000 32000 39000 46000 53000 60000 39MPa and G rockfill =11.7MPa for point 2. ref Shear modulus of core Gref [kPa] One can see that the curves associated to the solution set almost coincide with the optimum solution whereas the curve Figure 6: Solution of 9th population obtained from parameter set outside the identified solutions is irrelevant. It can be concluded, that even though the genetic algorithm does not have identified the exact optimum parameter set of this problem, this optimization technique is

5

TH TABLE 2: SET OF SOLUTIONS IN 9 POPULATION objective function in equation (1) equals to Ferr =0.0276 (with WKHSDUDPHWHUVİ DQGĮ  The outcome of the optimization core rockfill against the real dam case is shown in figure (8). The agreement G ref G ref Value of Point number error function between measured inclinometer values and optimized values of (MPa) (MPa) horizontal displacements might be interpreted as decent in this 1 31 14.5 1.06·10-03 real case. The Mohr-Coulomb model is for this case able to simulate the trend of horizontal displacements. Nevertheless, -03 2 39 11.7 1.06·10 the exact evolution of horizontal displacement is not perfectly 3 29 14.5 1.17·10-03 capture, optimization can still be improved. In fact, the Mohr- Coulomb model is a well-known model based on well-known

geotechnical parameters, but it is quite simple and then can

only be used as a first approximation. A better agreement

should be obtained with a more sophisticated constitutive 45 model. That’s what we expect by using in future, some of the other models of the Plaxis model library. 40

35 45 [m] [m] H 30 40

25 Modeled inclinometer values, synthetic case 35 [m] [m] 20 Point 1, fig (6) H 30

The height of the dam 15 Point 2, fig (6) 25 Inclinometer values

10 Point in intial population 20 Optimized values 5 0.01 0.03 0.05 0.07 0.09 0.11 0.13 0.15 of The height the dam 15 Horizontal displacement U [m] x

10 Figure 7: Synthetic case — horizontal displacement vs dam height 5 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18

Horizontal displacement Ux [m] Optimization results of the real dam case Figure 8: Measured and computed horizontal displacements An inverse analysis is here carried out against the real at different dam heights inclinometer measurements performed on the dam in figure (1). As in the optimization against the synthetic case above, the shear modulus Gref of the core and the shear modulus Gref of the rockfill were chosen as optimization variables. The search ௖௢௥௘ domains were also kept as: 29MPa ” ܩ௥௘௙ ” 93MPa and rockfill 12MPa ” G ref ” 56MPa. The result of this optimization gives the following parameter ௖௢௥௘ rockfill values: ܩ௥௘௙ = 31MPa, G ref =15MPa with the value of the

6

Concluding remarks horizontal displacements from inclinometer measurements were used. Identification might be improved if more measurements This research focused on validating the technique of inverse as well as measurements of different kind are included in the analysis, adopted to identify parameter values of constitutive objective function. models in an earth and rockfill dam application. As the search If inverse analysis applications become available in method in the optimization, the genetic algorithm was utilized. commercial finite element software, it would certainly be a Very good agreement was obtained in the optimization valuable tool for dam engineers assessing dam performance and against the synthetic case in this study. The synthetic case is a dam safety. good test for the mathematical basis used in the optimization, i.e. the objective function and the search algorithm. It can be Acknowledgements concluded, from this synthetic case study, that the technique of inverse analysis could be effectively used in this kind of earth The research presented was carried out as a part of "Swedish and rockfill dam application. However, in the synthetic case Hydropower Centre - SVC". SVC has been established by the presented here, the values of the inclinometer measurements Swedish Energy Agency, Elforsk and Svenska Kraftnät were simulated. Thus, the field measurements were not together with Luleå University of Technology, The Royal subjected to errors in the experimental result; something which Institute of Technology, Chalmers University of Technology are likely in practice. Perturbations on the field measurements and Uppsala University. www.svc.nu. The dam owner who might influence the effectiveness of the search method. The provided us with measured results from the dam studied is to be authors’ have planned, in future research, to examine how the acknowledged for all valuable assistance during the work. All genetic algorithm performs when the simulated experimental the members of the reference group for the project set by SVC result, in a synthetic case, is subjected to randomly generated are to be acknowledged for valuable contributions. Mr. Åke perturbations of different magnitude. Nilsson, WSP consulting AB is acknowledged for his interest A decent agreement was obtained in the optimization against in the project and valuable comments from a practical point of the real inclinometer measurements. The Mohr-Coulomb model view. Luleå University of Technology is finally to be captured the trend of increased horizontal displacements the acknowledged for providing possibilities for carry out the higher on the dam they were measured. The simulated research project. magnitudes of horizontal displacements did not agree very well with inclinometer measurements. Probably it is possible to get a References better agreement by choosing some other constitutive model (or models) in Plaxis. It is important to realize that a good [1] Duncan J.M. (1996). State of the art: limit equilibrium and finite-element agreement cannot be expected in an inverse analysis if the analysis of slopes. Journal of Geotechnical Engineering vol. 122, no. 7, constitutive models chosen are not able to simulate the real soil pp. 577-596. [2] Kovacevic N. (1994). Numerical analyses of rockfill dams, cut slopes behavior sufficiently accurate. So, this example should not be and road embankments. PhD Thesis, London University (Imperial viewed as an appropriate test of the mathematical basis utilized College of Science and Medicine), Faculty of Engineering. for optimization. Inverse analysis will presumably not work [3] Potts D.M., =GUDYNRYLü L. (2001). Finite element analysis in well if there are too many errors in the field measurements geotechnical engineering: application. vol. 2. Thomas Telford, London. [4] Muir Wood D. (2004). Geotechnical modelling. vol. 1. Taylor & Francis. either. Oxfordshire. In this study, only two optimization variables were utilized. [5] Gioda G., Sakurai S. (1987). Back analysis procedures for the The authors’ intention is to evaluate the mathematical basis for interpretation of field measurements in geomechanics. International a gradually increased number of optimization variables in the journal for Numerical and Analytical Methods in Geomechanics. vol. 11, no. 6, pp. 555-583. future. Anyway, inverse analysis with two optimization [6] Swoboda G., Ichikawa Y., Dong Q., Zaki M. (1999), Back analysis of variables is already time consuming, because of the complexity large geotechnical models. International journal for Numerical and of the problem. As an example, the 9th population in the Analytical Methods in Geomechanics. vol. 23, no. 13, pp. 1455-1472. optimization against the synthetic case was reached after 28 [7] Calvello M., Finno RJ. (2004). Selecting parameters to optimize in model calibration by inverse analysis. Computers and Geotechnics. vol. hours on a PC with 2.66GHz (2 processors- 4 cores) and 24GB 31. pp. 410-425. RAM. It might be possible to modify the search algorithm to [8] Finno RJ., Calvello M. (2005). Supported excavations: observational make it faster. method and inverse modeling. Journal of Geotechnical and It would be interesting to test the mathematical basis against Geoenvironmental Engineering. Vol.7. pp. 826-836. [9] Gens A., Ledesma A., Alonso EE. (1996). Estimation of parameters in field measurements of different kind also, like: deformations, geotechnical back analysis. Application to a tunnel excavation problem. pore pressures, total stresses, seepage etc. In this research, only Computers and Geotechnics vol. 18. pp. 1-27.

7

[10] Lecampion B., Contantinescu A. (2005). Sensitivity analysis for parameter identification in quasi-static poroelasticity. International Journal for Numerical and Analytical Methods in Geomechanics. vol. 29. pp. 163-185. [11] Lecampion B., Contantinescu A., Nguyen Minh D. (2002). Parameter identification for lined tunnels in viscoplastic medium. International journal for Numerical and Analytical Methods in Geomechanics. vol.26. pp.1191-1211. [12] Zentar R., Hicher P.Y., Moulin G. (2001). Identification of soil parameters by inverse analysis. Computers and Geotechnics. vol.28. pp.129-144. [13] Levasseur S., Malecot Y., Boulon M., Falvigny E. (2010). Statistical inverse analysis based on genetic algorithm and principal component analysis: Applications to excavation problems and pressuremeter tests. International Journal for Numerical and Analytical Methods in Geomechanics. vol.34.pp.471-491 [14] Gioda G., Maier G. (1980). Direct search solution of an inverse problem in elastoplasticity: identification of cohesion, friction angle and in situ stress by pressure tunnel tests. International Journal for Numerical Methods in Engineering. vol. 15, no. 12, pp. 1823-1848. [15] Tarantola A. (1987). Inverse problem theory. Elsevier. Amsterdam. [16] Yu Y., Zhang B., Yuan H. (2007). An intelligent displacement back- analysis method for earth-rockfill dams. Computers and Geotechnics. vol. 34, no. 6, pp. 423-434. [17] Pal S., Wije Wathugala G., Kundu S. (1996). Calibration of a constitutive model using genetic algorithms. Computers and Geotechnics. vol. 19, no. 4, pp. 325-348. [18] Gallagher K., Sambridge M.S., Drijkoningen G. (1991). Genetic algorithms-an evolution from Monte-Carlo methods for strongly non- linear geophysical optimization problems.Geophysical Research Letters. vol. 18, no. 12, pp. 2177-2180. [19] Haupt R.L., Haupt S.E. (2004). Practical genetic algorithms. Wiley: New York. [20] Levasseur S., Malécot Y., Boulon M., Flavigny E. (2007). Soil parameter identification using a genetic algorithm. International journal for Numerical and Analytical Methods in Geomechanics. vol. 32, no. 2, pp. 189-213. [21] Levasseur S., Malécot Y., Boulon M., Flavigny E. (2009). Statistical inverse analysis based on genetic algorithm and principal component analysis: method and developments using synthetic data. International journal for Numerical and Analytical Methods in Geomechanics. vol. 33, no. 12, pp. 1485-1511. [22] Rechea C., Levasseur S., Finno R. (2008). Inverse analysis techniques for parameter identification in simulation of excavation support systems. Computers and Geotechnics. vol. 35, no. 3, pp. 331-345. [23] PlAXIS. (2011). Finite Element Code for Soil and Rock Analysis. Version 2011-2. http://plaxis.nl [24] Reséndiz D., Romo M.P. (1972). Analysis of embankment deformation. ASCE. Vol.1. pp. 817-836. [25] Clough R.W., Woodward R.J. (1967). Analysis of embankment stresses and deformations. Journal of Soil Mechanics & Foundations Div. Vol 93, pp 529-549. [26] Bowles J.E. (1988). Foundation analysis and design. MacGraw-Hill. Berkshire. England. [27] Holland J.H. (1975). Adaptation in natural and artificial systems: an introductory analysis with applications to biology, control, and artificial intelligence. Cambridge: MIT press.

8 Paper B

Inverse Mohr-Coulomb soil parameter identification of an earth and rockfill dam by genetic algorithm optimization

Vahdati, P., Levasseur, S., Mattsson, H. and Knutsson, S

Electronic journal of geotechnical engineering, vol. 18, pp.5419-5440

,QYHUVH0RKU&RXORPE6RLO3DUDPHWHU ,GHQWLILFDWLRQRIDQ(DUWKDQG5RFNILOO 'DPE\*HQHWLF$OJRULWKP2SWLPL]DWLRQ 

Pooya Vahdati  'HSDUWPHQWRI&LYLO(QYLURQPHQWDODQG1DWXUDOUHVRXUFHVHQJLQHHULQJ/XOHn 8QLYHUVLW\RI7HFKQRORJ\7HO6(±/XOHn6ZHGHQ HPDLO3RR\D9DKGDWL#OWXVH

Séverine Levasseur 'RFWRU 'pSDUWHPHQW$U*(Q&R6HUYLFHGH*pRPpFDQLTXHHWJpRORJLHGHO LQJpQLHXU 8QLYHUVLWpGH/LqJH7HO/LqJH%HOJLXP HPDLOVHYHULQHOHYDVVHXU#XOJDFEH

Hans Mattsson $VVLVWDQW3URIHVVRU 'HSDUWPHQWRI&LYLO(QYLURQPHQWDODQG1DWXUDOUHVRXUFHVHQJLQHHULQJ/XOHn 8QLYHUVLW\RI7HFKQRORJ\7HO6(±/XOHn6ZHGHQ HPDLO+DQV0DWWVVRQ#OWXVH

Sven Knutsson 3URIHVVRU 'HSDUWPHQWRI&LYLO(QYLURQPHQWDODQG1DWXUDOUHVRXUFHVHQJLQHHULQJ/XOHn 8QLYHUVLW\RI7HFKQRORJ\7HO6(±/XOHn6ZHGHQ HPDLO6YHQ.QXWVVRQ#OWXVH

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

-5419 -  Vol. 18 [2013], Bund. X 5420

INTRODUCTION 7KH ILQLWH HOHPHQW PHWKRG )(0  KDV EHHQ ZLGHO\ XVHG WR DQDO\]HHDUWKDQGURFNILOOGDPV 'XQFDQ.RYDFHYLF3RWWVDQG=GUDYNRYLü0XLU:RRG 7KHILQLWHHOHPHQW PHWKRG ZDV LQWURGXFHG WR WKH JHRWHFKQLFDO HQJLQHHULQJ SURIHVVLRQ E\ &ORXJK DQG :RRGZDUG  RQWKHVWDELOLW\DQGSHUIRUPDQFHRIVORSHVDQGHPEDQNPHQWV,QDILQLWHHOHPHQWDQDO\VLVD SURSHU FRQVWLWXWLYH PRGHO KDV WR EH FKRVHQ IRU HDFK SDUW RI WKHGDPLQRUGHUWRVLPXODWHWKH UHODWLRQEHWZHHQVWUHVVHVDQGVWUDLQV7KHYDULRXV]RQHVRIDQHDUWKDQGURFNILOOGDPDUHQRUPDOO\ FRQVWUXFWHGRIGLIIHUHQWVRLOPDWHULDOVIRUZKLFKWKHVWUHVVVWUDLQUHVSRQVHFRXOGYDU\FRQVLGHUDEO\ ,Q DGGLWLRQ VXLWDEOH YDOXHV KDYH WR EH DVVLJQHG WR WKH SDUDPHWHUV LQFOXGHG LQ WKH FRQVWLWXWLYH PRGHOFKRVHQIRUHDFKGDPSDUWWREHDEOHWRFDSWXUHWKHLPSRUWDQWIHDWXUHVRIVRLOEHKDYLRU,Q JHQHUDOH[SHULPHQWDOGDWD ODERUDWRU\DQGRUILHOGWHVWV RIHDFKVRLODUHQHHGHGDVDEDVLVIRUWKLV SDUDPHWHU HYDOXDWLRQ +RZHYHU PDQ\ GDPV DUH ROG DQG OLPLWHG LQIRUPDWLRQ PLJKW EH DYDLODEOH UHJDUGLQJWKHVRLOPDWHULDOVEHLQJXVHGLQWKHGDPVWUXFWXUHV,QGDPVLWLVQRUPDOO\YHU\GLIILFXOW WRWDNHXSVRLOVDPSOHVIRUWHVWLQJHVSHFLDOO\IURPWKHFHQWUDOLPSHUYLRXVSDUWVLQFHWKLVPLJKW DIIHFW WKH GDP SHUIRUPDQFH DQG WKH VDIHW\ RI WKH GDP )RU GDPV LW ZRXOG EH DGYDQWDJHRXV LI FRQVWLWXWLYHSDUDPHWHUYDOXHVFRXOGEHGHWHUPLQHGZLWKVRPHQRQGHVWUXFWLYHPHWKRG ,QYHUVH DQDO\VLV SURYLGHV D SRVVLELOLW\ WR GHWHUPLQH WKH FRQVWLWXWLYH EHKDYLRU RI GLIIHUHQW PDWHULDOV ZLWKLQ WKH GDP VWUXFWXUH XQGHU WKH FRQGLWLRQ WKDW WKH GDP KDV EHHQ HTXLSSHG ZLWK YDULRXV LQVWUXPHQWDWLRQV IRU PRQLWRULQJ GDP SHUIRUPDQFH ZKLFK UHFRUG GDWD VXFK DV SRUH SUHVVXUHV GHIRUPDWLRQV WRWDO VWUHVVHV DQG VHHSDJH HWF ,Q WKH SURSRVHG LQYHUVH DQDO\VLV DOVR QDPHGEDFNDQDO\VLVDILQLWHHOHPHQWPRGHOLVFDOLEUDWHGDXWRPDWLFDOO\E\FKDQJLQJWKHYDOXHVRI WKHLQSXWSDUDPHWHUVRIWKHVHOHFWHGFRQVWLWXWLYHPRGHOLQGLIIHUHQWGDP]RQHVXQWLOWKHGLVFUHSDQF\ EHWZHHQ PHDVXUHG UHVXOWV E\ GDP LQVWUXPHQWDWLRQV DQG WKH FRUUHVSRQGLQJ FRPSXWHG UHVXOWV LV PLQLPL]HG7KHILQLWHHOHPHQWPRGHOFDOLEUDWLRQLVFRQGXFWHGE\DSSO\LQJDQRSWLPL]DWLRQPHWKRG WRVHDUFKIRUWKHEHVWSRVVLEOHVROXWLRQV+RZHYHUWKHUHDUHVRPHIDFWRUVZKLFKVWURQJO\DIIHFW SURSHUFDOLEUDWLRQVXFKDVWKHQXPEHURISDUDPHWHUVRSWLPL]HGZKLFKGHSHQGVRQWKHQXPEHURI SDUDPHWHUVLQWKHFRQVWLWXWLYHPRGHOFKRVHQIRUHDFKGDP]RQHDQGWKHW\SHDQGDPRXQWRIGDP PRQLWRULQJPHDVXUHPHQWVXVHGIRUFDOLEUDWLRQ7KHLQYHUVHDQDO\VLVWHFKQLTXHZDVLQWURGXFHGWR WKH JHRWHFKQLFDO ILHOG E\ *LRGD DQG 6DNXUDL   IRU WKH SXUSRVH RI LGHQWLILFDWLRQ RI HODVWLF PDWHULDOSURSHUWLHVRILQVLWXURFNPDVVHV7KHLQYHUVHDQDO\VLVPHWKRGKDVVLQFHWKHQEHHQDSSOLHG LQPDQ\GLIIHUHQWW\SHVRIJHRWHFKQLFDODSSOLFDWLRQV 6ZRERGDHWDO*LRGDDQG0DLHU 7DUDQWROD  &DOYHOOR DQG )LQQR  )LQQR DQG &DOYHOOR  /HGHVPD HW DO  /HFDPSLRQHWDO/HFDPSLRQHWDO=HQWDUHWDO/HYDVVHXUHWDO

WHVWVWXGLHVVHHHJ/HYDVVHXUHWDO  5HFKHDHWDO  3DSRQHWDO   +RZHYHU WR WKH DXWKRUV¶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

 Figure 1: &URVVVHFWLRQRIWKHGDP Vol. 18 [2013], Bund. X 5422

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pQGL] DQG 5RPR  &ORXJK DQG :RRGZDUG  7KHH[FHVVSRUHSUHVVXUHVLQWKHFRUHZHUHGLVVLSDWHGE\FRQVROLGDWLRQDIWHUFRQVWUXFWLRQRI WKHGDP7KH0RKU&RXORPEHODVWLFSHUIHFWO\SODVWLFPRGHOZDVFKRVHQIRUDOOGDP]RQHVDVDQ LQLWLDOFKRLFHIRUWKLVUHVHDUFKFKLHIO\EHFDXVHRILWVVLPSOLFLW\7DEOH  VXPPDUL]HVWKH0RKU

&RXORPEPRGHOSDUDPHWHUVZKLFKDUHDSSOLHGIRUWKLVFDVHVWXG\7KHSDUDPHWHUYDOXHVRIJ X J V  ij N[ DQGN\IRUDOOVRLOOD\HUVLQWKHGDPZHUHSURYLGHGE\WKHGDPRZQHU$SSURSULDWHYDOXHVRI Ȟ DQGF ZHUHFKRVHQEDVHGRQDGYLFHLQWKHUHIHUHQFH%RZOHV  $OOWKHGLODWDQF\DQJOHV ZHUHFDOFXODWHGIURPWKHVDPHHPSLULFDOUHODWLRQFRPPRQO\DFFHSWHGIRUGHQVHVDQGȥ  ij ƒ

DV SURSRVHG E\ 3OD[LV   $ KLJK YDOXH RI WKH VKHDU PRGXOXV *UHIZDVFRQVLGHUHGIRUWKH IRXQGDWLRQWRPDNHLWULJLG+RZHYHUQRUPDOO\WKHVWLIIQHVVRIVRLOLVDVVXPHGWRLQFUHDVHOLQHDUO\

ZLWKGHSWKZLWKUHJDUGWRWKHZHLJKWRIVRLODVKHDUPRGXOXVLQFUHPHQWSHUXQLWRIGHSWK*LQFZDV GHILQHGIRUWKHFRUHILOWHUVDQGWKHURFNILOOPDWHULDOV7KLVDSSURDFKVKRXOGSURYLGHPRUHUHDOLVWLF

GHIRUPDWLRQV ZLWKLQ WKH GDP VWUXFWXUH 7KH UHIHUHQFH YDOXHV RI VKHDU PRGXOXV *UHIDQGVKHDU

PRGXOXV LQFUHPHQW *LQF RI WKH ILOWHU PDWHULDOV ZHUH SRVVLEOH WR LGHQWLI\ DW DQ HDUO\ VWDJH E\ FDOLEUDWLRQDJDLQVWLQFOLQRPHWHUUHVXOWVVLQFHWKHILOWHUVLQIOXHQFHKRUL]RQWDOGHIRUPDWLRQVOHVVWKDQ

WKHFRUHDQGURFNILOOPDWHULDO7KHYDOXHVRI*LQFRIWKHFRUHDQGWKHURFNILOOZHUHREWDLQHGIURP

RSWLPL]DWLRQDJDLQVWWKHUHDOGDPFDVHLQWKLVVWXG\7KH*UHIYDOXHVRIFRUHDQGURFNILOOOD\HUV ZHUHFKRVHQDVUHSUHVHQWDWLYHORZPDJQLWXGHYDOXHVEDVHGRQDVHQVLWLYLW\DQDO\VLV 7KHGDPFRUHZDVPRGHOHGDVXQGUDLQHGFDVH % LQWKH3OD[LVVRIWZDUH7KHXQGUDLQHGVKHDU VWUHQJWKQHHGHGLQFDVH % ZDVDVVXPHGWRLQFUHDVHOLQHDUO\ZLWKGHSWKDQGZDVFDOFXODWHGZLWK WKHHIIHFWLYHSDUDPHWHUVij DQGܿᇱ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ol. 18 [2013], Bund. X 5423

FRQGLWLRQ7KHILUVWEHUPZDVFRQVWUXFWHGEHIRUHULVLQJXSWKHUHVHUYRLUDQGWKHVHFRQGEHUPZDV FRQVWUXFWHGZKHQWKHUHVHUYRLUZDWHUOHYHOZDVLQLWVKLJKHVWOHYHO7KHH[FHVVSRUHSUHVVXUHVLQ WKHFRUHZHUHGLVVLSDWHGDIWHUWKHUHVHUYRLUUHDFKHGLWVKLJKHVWOHYHOEXWQRWDIWHUFRQVWUXFWLRQRI WKHVHFRQGEHUP

 Figure 2:)LQLWHHOHPHQWPRGHORIWKHFURVVVHFWLRQRIWKHGDP

Table 1: 0DWHULDOSURSHUWLHVRIWKHURFNILOOGDP

\N[N ׋ F LQFȞ UHI* ȖXȖV * 'DP]RQHV N1P 03D N3D  N3D ƒ  PV &RUH        ā )LQHILOWHU        ā &RDUVHILOWHU        ā 5RFNILOO        ā %HUPV        ā )RXQGDWLRQ        ā

1RWH J X LVWKHXQLWZHLJKWDERYHWKHSKUHDWLFOHYHOJ V LVWKHXQLWZHLJKWEHORZWKHSKUHDWLF OHYHO*'UHI LVWKHVKHDUPRGXOXV*'LQFLVWKHVKHDUPRGXOXVLQFUHPHQWȞ LVWKHHIIHFWLYH3RLVVRQ¶V UDWLRF LVWKHHIIHFWLYHFRKHVLRQij LVWKHHIIHFWLYHIULFWLRQDQJOHDQGN[ DQGN\DUHWKHK\GUDXOLF FRQGXFWLYLW\LQKRUL]RQWDODQGYHUWLFDOGLUHFWLRQUHVSHFWLYHO\

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ol. 18 [2013], Bund. X 5424

ERROR FUNCTION ,Q RUGHU WRGHILQH WKH GLIIHUHQFH EHWZHHQ WKH PHDVXUHG YDOXHV DQG ILQLWH HOHPHQWFRPSXWHG YDOXHVLWLVHVVHQWLDOWRGHILQHDSURSHUHUURUIXQFWLRQ,QWKLVUHVHDUFKIROORZLQJ/HYDVVHXUHWDO

 DVFDODUHUURUIXQFWLRQ)HUUEDVHGRQWKHOHDVWVTXDUHPHWKRGZDVLQWURGXFHG

ൗ 1    ൫8HL8QL൯  )HUU  ቌ  ෍  ቍ  1 ǻ8    L  L

WKDWGHWHUPLQHVWKHGLIIHUHQFHEHWZHHQH[SHULPHQWDOO\PHDVXUHGYDOXHV8HLDQGQXPHULFDOO\

FRPSXWHG YDOXHV 8QL  IURP WKH 1 PHDVXULQJ SRLQWV 7KH WHUP ǻ8L JLYHV WKH ZHLJKW RI WKH

GLVFUHSDQF\ EHWZHHQ 8HL DQG 8QL 7KH SDUDPHWHU ǻ8L UHODWHV WR WKH QXPHULFDO DQG H[SHULPHQWDO XQFHUWDLQWLHVDWWKHPHDVXUHPHQWSRLQWLGHILQHGDV

߂8L İĮ8HL  

ZKHUHWKHSDUDPHWHUİLVDQDEVROXWHHUURURIPHDVXUHPHQWVDQGSDUDPHWHUĮLVDGLPHQVLRQOHVV UHODWLYHHUURURIPHDVXUHPHQWV

SEARCH ALGORITHM 7KH JHQHWLF DOJRULWKP EHORQJV WR WKH VXEFDWHJRU\ RI VWRFKDVWLF VHDUFK PHWKRGV DQG ZDV GHYHORSHGE\-RKQ+ROODQGLQWKHVDWWKH8QLYHUVLW\RI0LFKLJDQWKHPHWKRGZDVLQVSLUHG E\'DUZLQ¶VWKHRU\RIHYROXWLRQ7KHJHQHWLFDOJRULWKPLVDJOREDORSWLPL]DWLRQWHFKQLTXHZKLFKLV EDVHGRQUDQGRPL]HGRSHUDWRUVVXFKDVVHOHFWLRQFURVVRYHUDQGPXWDWLRQ7KHVHDUFKPHWKRGFDQ DSSURDFKDVHWRIVROXWLRQVFORVHWRWKHRSWLPXPVROXWLRQZLWKRXWEHLQJWUDSSHGLQWRORFDORSWLPD LQWKHJLYHQVHDUFKGRPDLQ7KHJHQHWLFDOJRULWKPLVEULHIO\GHVFULEHGEHORZ$PRUHWKRURXJK GHVFULSWLRQFDQHJEHIRXQGLQWKHERRNE\+DXSW  7KHRSWLPL]DWLRQSURJUDPWKDWLVXVHG LQWKLVUHVHDUFKZDVGHYHORSHGLQWKH)2575$1ODQJXDJHE\/HYDVVHXUHWDO DQG  IRUJHRWHFKQLFDOVWXGLHV

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¶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ol. 18 [2013], Bund. X 5425

FUHDWHG1RZWKHQHZSRSXODWLRQFRQVLVWLQJRISDUHQWVDQGFKLOGUHQLVFRPSOHWHG0XWDWLRQVDSSO\ IRUSUHYHQWLQJWKHJHQHWLFDOJRULWKPIURPFRQYHUJLQJWRRIDVWEHIRUHVZHHSLQJWKHHQWLUHVHDUFK GRPDLQ 7KHUHIRUH VRPH RI WKH LQGLYLGXDOV DUH UDQGRPO\ VHOHFWHG WR IOLS WKHLU FKDUDFWHU VHH )LJXUH

 Figure 3:5HSURGXFWLRQRIRIIVSULQJ FKLOGUHQ IURPWKHSDLURISDUHQWV 

 Figure 4: 0XWDWLRQRIDQLQGLYLGXDO 

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

LQFUHPHQW*LQFRIWKHFRUHDQGWKHURFNILOOPDWHULDOVZHUHWZRRIWKHPRVWVHQVLWLYHSDUDPHWHUVIRU

WKHKRUL]RQWDOGHIRUPDWLRQEHKDYLRULQWKHGRZQVWUHDP]RQHVRIWKHGDP7KHUHIRUH*LQF RIWKH FRUHDQGWKHURFNILOOZHUHFRQVLGHUHGDVWKHWZRRSWLPL]DWLRQYDULDEOHVLQWKLVVWXG\7KHVHDUFK FRUH GRPDLQZDVOLPLWHGE\WKHVDPHORZHUDQGXSSHUERXQGVIRUWKHVHWZRYDULDEOHVDVN3D” *LQF URFNILOO FRUH  *LQF ”N3D7KHRSWLPXPVROXWLRQLQWKLVVWXG\LVNQRZQDVWKHYDOXHV *LQF  N3DDQG URFNILOO *LQF  N3D,QWKLVUHVHDUFKHDFKRSWLPL]DWLRQSDUDPHWHULVHQFRGHGE\E\WHVLQWKHJHQHWLF DOJRULWKP7KHQWKHVHDUFKGRPDLQFRQVLVWVRI WLPHVQRGHV6LQFHWKHVDPHERXQGFRQVWUDLQWV FRUH DUHDSSOLHGIRUERWKRSWLPL]DWLRQYDULDEOHVWKHPHVKVL]HIRUWKHVHDUFKVSDFHLVHTXDOWRǻ *LQF   URFNILOO ǻ *LQF N3D 7KHRSWLPL]DWLRQUHVXOWVRIWKHV\QWKHWLFFDVHDUHVKRZQLQILJXUHV    DQG  ,QILJXUH  WKHLQLWLDOSRSXODWLRQJHQHUDWHGE\WKHJHQHWLFDOJRULWKPLVSUHVHQWHG7KHLQLWLDOSRSXODWLRQ Vol. 18 [2013], Bund. X 5426

UHSUHVHQWV WKH VWDUW RI WKH VHDUFK IRU WKH RSWLPXP DQG FRQVLVWV RI  SRLQWV LQ WKH WZR GLPHQVLRQDOVHDUFKVSDFHVSDQQHGE\WKHRSWLPL]DWLRQYDULDEOHVDQGWKHLUERXQGDULHV )RUHDFKRIWKHIROORZLQJSRSXODWLRQVKDOIRIWKHQXPEHURISRLQWVLQWKHLQLWLDOSRSXODWLRQLV DSSOLHG SRLQWV 7KHSRSXODWLRQVL]HLV FKRVHQ EDVHGRQDVWXG\SUHVHQWHGLQDVXEVHTXHQW FKDSWHU7KHHUURUIXQFWLRQGHILQHGLQ(TXDWLRQLVFRQVLGHUHGZLWKWKHSDUDPHWHUVİ DQGĮ  7KDWLVWRVD\WKHIXQFWLRQJLYHQE\(TXDWLRQLVUHGXFHGWR

ൗ 1    )   ቌ  ෍ ൫8 8 ൯ ቍ    HUU 1 HL QL L 

DQGKDVWKHXQLWRIOHQJWK $IWHU ILYH LWHUDWLRQV WKH JHQHWLF DOJRULWKP FRQYHUJHG WR WKH WKUHH SDUDPHWHU VHWV JLYHQ LQ 7DEOH)LJXUHVKRZVWKDWRQO\WKHEHVWVROXWLRQLGHQWLILHGLVYHU\FORVHWRWKHRSWLPDORQH GHVSLWHWKHIDFWWKDWDOOVROXWLRQVSUHVHQWTXLWHVLPLODUHUURUIXQFWLRQYDOXHVDQGWKH\FDQDOOZHOO UHSURGXFHKRUL]RQWDOGLVSODFHPHQWYHUVXVKHLJKWRIWKHGDPVHH)LJXUH7KLVUHVXOWLVSUREDEO\ GXHWRWKHSDUWLFXODUWRSRORJ\RIWKHHUURUIXQFWLRQDVZLOOEHVHHQLQWKHQH[WVXEVHFWLRQ)URP WKLV VHW RI VROXWLRQV WKH XVHU FDQ FKRRVH D VROXWLRQ WKDW ³IHHOV FRUUHFW´ IURP DQ HQJLQHHULQJ SHUVSHFWLYHDQGKDVDORZYDOXHRIWKHHUURUIXQFWLRQ

,QLWLDOSRSXODWLRQ 2SWLPXPVROXWLRQ 

















Shear modulus increment of core [kPa] modulus Shear 



             Shear modulus increment of rockfill [kPa]  Figure 5: ,QLWLDOSRSXODWLRQ Vol. 18 [2013], Bund. X 5427

6ROXWLRQVWKSRSXODWLRQ 2SWLPXPVROXWLRQ

 1    2 3      

Shear modulus increment of core [kPa] modulus Shear     Shear modulus increment of rockfill [kPa]  Figure 6:6ROXWLRQVRIWKSRSXODWLRQDQGRSWLPXPVROXWLRQ   Table 2:6HWRIVROXWLRQVLQWKSRSXODWLRQ

FRUH URFNILOO *LQF *LQF YDOXHRIHUURU SRLQWQXPEHU N3D N3D IXQFWLRQ P     ā    ā    ā 7RJRIXUWKHULQWKHDQDO\VLVDQGWRKDYHDQLGHDRIWKHFDSDELOLWLHVRIWKHJHQHWLFDOJRULWKP ZHFDQFRPSDUHGLIIHUHQWQXPHULFDOUHVXOWV,Q)LJXUHKRUL]RQWDOGLVSODFHPHQWYHUVXVWKHKHLJKW RIWKHGDPLVVKRZQIRUWKHLQFOLQRPHWHUSRVLWLRQLQWKHGDP7KHLQFOLQRPHWHUYDOXHVIURPWKH V\QWKHWLF FDVH UHSUHVHQW WKH RSWLPXP VROXWLRQ ,Q WKLV ILJXUH QXPHULFDO FXUYHV REWDLQHG IURP FRUH VHYHUDOSDUDPHWHUVHWVDUHFRPSDUHG)LUVWZHFRQVLGHUDFXUYHREWDLQHGIURP *LQF N3DDQG URFNILOO  *LQF N3DLQWKHLQLWLDOSRSXODWLRQIRUZKLFKWKHHUURUIXQFWLRQYDOXHLVHTXDOWRā  7KHUHDIWHUZHFRQVLGHUFXUYHVREWDLQHGIURPWKHLGHQWLILHGVROXWLRQVHWHJSRLQWVDQGZKLFK JDYHWKHVPDOOHVWHUURUIXQFWLRQYDOXHVVHH7DEOH 2QHFDQVHHWKDWWKHFXUYHVDVVRFLDWHGWRWKHVROXWLRQVHWDOPRVWFRLQFLGHZLWKWKHRSWLPXP VROXWLRQ ZKHUHDV WKH FXUYH REWDLQHG IURP D SDUDPHWHU VHW RXWVLGH WKH LGHQWLILHG VROXWLRQV LV LUUHOHYDQW  Vol. 18 [2013], Bund. X 5428









 0RGHOHGLQFOLQRPHWHU YDOXHVV\QWKHWLFFDVH  3RLQWILJ



The height of the dam [m] 3RLQWILJ

 3RLQWLQLQWLDOSRSXODWLRQ



         

Horizontal displacement [m]  Figure 7: 6\QWKHWLFFDVH²KRUL]RQWDOGLVSODFHPHQWYHUVXVKHLJKWRIWKHGDP

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ǻ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ol. 18 [2013], Bund. X 5429

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

 Figure 87RSRORJ\RIHUURUIXQFWLRQ Vol. 18 [2013], Bund. X 5430

 Figure 97RSRORJ\RIHUURUIXQFWLRQ YLHZIURPDERYH  

 Figure 10: /HYHOFRQWRXUVRIHUURUIXQFWLRQ 

STUDY OF POPULATION SIZE ,QWKHJHQHWLFDOJRULWKPDSRSXODWLRQVL]HKDVWREHFKRVHQ7KLVFKRLFHLVYHU\LPSRUWDQWIRU WKHHIILFLHQF\RIWKHDOJRULWKP,WLVLQJHQHUDOQRWHDV\WRFKRRVHWKHSRSXODWLRQVL]HVLQFHWKH EHVWFKRLFHGHSHQGVRQWKHDSSOLFDWLRQLQZKLFKWKHJHQHWLFDOJRULWKPLVXVHG,IWKHSRSXODWLRQ VL]HLVWRRVPDOOWKHJHQHWLFDOJRULWKPLVPRUHOLNHO\WRFRQYHUJHTXLFNO\EXWLWPLJKWEHWRDORFDO Vol. 18 [2013], Bund. X 5431

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ഥHUUDERXWā WRā 7KHUHIRUHDQDYHUDJHHUURUIXQFWLRQYDOXH)ഥHUURIā FDQEHFRQVLGHUHGDVDQXSSHUOLPLWIRUDQDFFHSWDEOHVROXWLRQVHW$LQWKHJHQHWLFDOJRULWKP,Q WDEOH  VRPHVWDWLVWLFDOGDWDDUHSUHVHQWHGIRUWKHVROXWLRQVHWV$UHSUHVHQWHGDVSRLQWVLQILJXUH    7REHDEOHWRGHWHUPLQHWKHEHVWFKRLFHRIWKHSRSXODWLRQVL]HDQHZVHWRIVROXWLRQVZHUH GHILQHG QDPHG VROXWLRQ VHW % ,Q DQ LQYHUVH FRPSXWDWLRQ ZLWK WKH JHQHWLF DOJRULWKP D ORW RI SRWHQWLDOO\JRRGVROXWLRQVLGHQWLILHGLQWKHJHQHWLFDOJRULWKPPLJKWH[LVWDSDUWIURPWKHVROXWLRQ  VHW$6ROXWLRQVHW%LVFRPSRVHGRIDOOPRQLWRUHGSRLQWVZLWK )HUU ൑ā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ol. 18 [2013], Bund. X 5432

FRPSXWDWLRQRQWKLVV\QWKHWLFFDVHRQD3&ZLWKSURFHVVRUVFRUHV#*+]0%FDFKH DQG*%5$0 7KHFRPSXWDWLRQWLPHLVDQLPSRUWDQWLVVXHWRFRQVLGHU$ODUJHSDUWRIWKHILQLWH HOHPHQWFRPSXWDWLRQVGHDOVZLWKJHQHUDWLRQRIWKHLQLWLDOSRSXODWLRQ WZLFHWKHSRSXODWLRQVL]H  7KHHVWLPDWHGWLPHIRUJHQHUDWLQJWKHLQLWLDOSRSXODWLRQIRUHDFKSRSXODWLRQVL]HLVSUHVHQWHGLQ WDEOH   )RU WKH SRSXODWLRQ VL]HV  DQG  LW FDQ EH VHHQ WKDW WKH SRLQWVLQILJXUH  DUHUDWKHU   VSUHDGDQGPDQ\RIWKHPDUHDERYHWKHDFFHSWDEOHXSSHUOLPLWRI ഥ)HUU = ā  ,QWDEOH  DQG ILJXUH  LWLVHYLGHQWWKDWWKHSRSXODWLRQVL]HVDQGKDYHWKHORZHVWQXPEHURIFRPPRQ VROXWLRQVZLWKLQWKHLURZQDQGRWKHUSRSXODWLRQVL]HV7KHSRSXODWLRQVL]HVDQGDUHWRRVPDOO WR EH HIILFLHQW IRU WKH V\QWKHWLF FDVH VLQFH WKH YDOXHV RI )ഥHUU DUH WRR KLJK DQG HDFK LQYHUVH FRPSXWDWLRQGRHVQRWQHFHVVDULO\FRQYHUJHWRWKHVDPHVHWRIVROXWLRQV %DVHGRQWKHUHVXOWVLQWKLVVWXG\DOOWKHSRSXODWLRQVL]HVDQGFRXOGEHXVHG IRU WKH V\QWKHWLF FDVH ,Q JHQHUDO ODUJHU SRSXODWLRQ VL]HV VHHP WR JLYH EHWWHU SHUIRUPDQFH EXW ORQJHU FRPSXWDWLRQ WLPH 7KH RQO\ H[FHSWLRQ LV WKDW WKH SRSXODWLRQ VL]H  KDV RQH PRUH SRLQW  VDWLVI\LQJ)ഥHUU ൑ā FRPSDUHGWRWKHSRSXODWLRQVL]HVHHILJXUH  DQGWDEOH  7KDWLV FRQVLGHUHG DV D FRLQFLGHQFH $FFRUGLQJ WR ILJXUH   ILJXUH  WDEOH  DQGWDEOH  WKH   SRSXODWLRQ VL]HV  DQG  KDYH PRUH SRLQWV VDWLVI\LQJ  )ഥHUU ൑ ā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ol. 18 [2013], Bund. X 5433

3RSXODWLRQVL]H 3RSXODWLRQVL]H 3RSXODWLRQVL]H 3RSXODWLRQVL]H 3RSXODWLRQVL]H 3RSXODWLRQVL]H (ā

(ā

(ā

(ā

(ā

(ā

(ā

(ā Average error function value of the set of solutions A the set of solutions value of error function Average

(ā              Number of finite element computations  Figure 11:$YHUDJHHUURUIXQFWLRQYDOXHYHUVXVQXPEHURIILQLWHHOHPHQWFRPSXWDWLRQV  Table 3: 6WDWLVWLFDOGDWDIRUVROXWLRQVHWV$ SRLQWV LQILJXUH   3RSXODWLRQ $YHUDJHHUURU 6ROXWLRQVHWV$ZLWK ഥ  VL]H IXQFWLRQYDOXH )HUU ൑ā   ā

 ā

 ā

 ā

 ā

 ā   Vol. 18 [2013], Bund. X 5434

3RSXODWLRQVL]H 3RSXODWLRQVL]H 3RSXODWLRQVL]H 3RSXODWLRQVL]H 3RSXODWLRQVL]H













 Amount of common solutions 

 population population population population population size 150 size 120 size 90 size 75 size 45

Figure 12: &RPSDULVRQRIFRPPRQVROXWLRQVDPRQJGLIIHUHQWSRSXODWLRQVL]HV  Table 4: 6WDWLVWLFDOGDWDIRUVROXWLRQVHWV%

3HUFHQWRIWKH 3UREDELOLW\ XQLTXHVROXWLRQV 7RWDOQXPEHU RIILQGLQJ 1XPEHURI (VWLPDWHGWLPHIRU WKDWZHUHUHSHDWHG 3RSXODWLRQ RIVROXWLRQV RQHVROXWLRQ XQLTXHVROXWLRQV JHQHUDWLQJWKH DWOHDVWRQFH VL]H ZLWK DPRQJDOO ZLWKRXW LQLWLDOSRSXODWLRQ DPRQJWKHWRWDO PXOWLSOLFLW\ WKHRWKHU PXOWLSOLFLW\ KRXUV  QXPEHURI VROXWLRQV VROXWLRQV

                                     Table 5:$FORVHUORRNRQWKHVHYHQXQLTXHVROXWLRQVREWDLQHGZLWKSRSXODWLRQVL]H WKDWFRXOGQRWEHIRXQGZLWKSRSXODWLRQVL]H Population size 150/120

FRUH URFNILOO Solution number *LQF (kPa) *LQF (kPa) Error function value 204 111 2.94·10-3 1 153 111 2.94·10-3 162 128 2.64·10-3 2 153 120 2.65·10-3 69 153 2.81·10-3 3 61 153 2.89·10-3 474 61 1.33·10-3 4 525 61 1.16·10-3 221 128 2.31·10-3 5 162 128 2.64·10-3 457 69 2.25·10-4 6 449 86 2.65·10-4 407 94 8.64·10-4 7 331 86 9.19·10-4 

OPTIMIZATION RESULTS OF THE REAL DAM CASE $QLQYHUVHDQDO\VLVLVFDUULHGRXWKHUHDJDLQVWWKHUHDOLQFOLQRPHWHUPHDVXUHPHQWVSHUIRUPHG RQ WKH GDP LQ ILJXUH   $V LQ WKH RSWLPL]DWLRQ DJDLQVW WKH V\QWKHWLF FDVH DERYH WKH VKHDU PRGXOXVLQFUHPHQWRIWKHFRUHDQGWKHURFNILOOZHUHFKRVHQDVRSWLPL]DWLRQYDULDEOHV7KHVHDUFK FRUH URFNILOO GRPDLQZDVDOVRNHSWWKHVDPHIRUERWKSDUDPHWHUVN3D”*LQF ”N3DDQGN3D” *LQF ”N3D7KHSRSXODWLRQVL]HWKDWZDVHYDOXDWHGDVVXLWDEOHLQWKHSUHYLRXVFKDSWHUZDV XVHGLQWKHJHQHWLFDOJRULWKP $PRQJDOOSRVVLEOHVROXWLRQVLQWKLVRSWLPL]DWLRQWKHFKRVHQVROXWLRQZLWKWKHORZHVWHUURU FRUH URFNILOO IXQFWLRQYDOXHLQWKHVROXWLRQVHWJLYHVWKHSDUDPHWHUYDOXHV *LQF N3DDQG *LQF N3D ZLWKWKHYDOXHRIWKHHUURUIXQFWLRQLQHTXDWLRQ  HTXDOWR)HUU   ZLWKWKHSDUDPHWHUVİ  DQGĮ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

-5435 -  Vol. 18 [2013], Bund. X 5436











 ,QFOLQRPHWHUYDOXHV 2SWLPL]HGYDOXHV 

The height of the dam [m] 



           Horizontal displacement [m]  Figure 13:0HDVXUHGDQGFRPSXWHGKRUL]RQWDOGLVSODFHPHQWVDWGLIIHUHQWKHLJKWRIWKHGDP

CONCLUDING REMARKS 7KLV UHVHDUFK IRFXVHG RQ YDOLGDWLQJ WKH WHFKQLTXH RI LQYHUVH DQDO\VLV DGRSWHG WR LGHQWLI\ SDUDPHWHUYDOXHVRIFRQVWLWXWLYHPRGHOVLQDQHDUWKDQGURFNILOOGDPDSSOLFDWLRQ7KHFRQVWLWXWLYH PRGHO FKRVHQ LQ WKLV VWXG\ ZDV WKH 0RKU&RXORPE PRGHO $V WKH VHDUFK PHWKRG LQ WKH RSWLPL]DWLRQWKHJHQHWLFDOJRULWKPZDVXWLOL]HG 9HU\ JRRG DJUHHPHQW ZDV REWDLQHG LQ WKH RSWLPL]DWLRQ DJDLQVW WKH V\QWKHWLF FDVH LQ WKLV VWXG\7KHV\QWKHWLFFDVHLVDJRRGWHVWRIWKHPDWKHPDWLFDOEDVLVXVHGLQWKHRSWLPL]DWLRQLHWKH HUURUIXQFWLRQDQGWKHVHDUFKDOJRULWKP,WFDQEHFRQFOXGHGIURPWKLVV\QWKHWLFFDVHVWXG\WKDW WKHWHFKQLTXHRILQYHUVHDQDO\VLVFRXOGEHHIIHFWLYHO\XVHGLQWKLVNLQGRIHDUWKDQGURFNILOOGDP DSSOLFDWLRQ +RZHYHU LQ WKH V\QWKHWLF FDVH SUHVHQWHG KHUH WKH YDOXHV RI WKH LQFOLQRPHWHU PHDVXUHPHQWVZHUHVLPXODWHG7KXVWKHILHOGPHDVXUHPHQWVZHUHQRWVXEMHFWHGWRHUURUVLQWKH H[SHULPHQWDOUHVXOWVRPHWKLQJWKDWLVOLNHO\LQSUDFWLFH3HUWXUEDWLRQVRQWKHILHOGPHDVXUHPHQWV PLJKW LQIOXHQFH WKH HIIHFWLYHQHVV RI WKH VHDUFK PHWKRG 7KH DXWKRUV KDYH SODQQHG LQ IXWXUH UHVHDUFKWRH[DPLQHKRZWKHJHQHWLFDOJRULWKPSHUIRUPVZKHQWKHVLPXODWHGH[SHULPHQWDOUHVXOW LQ D V\QWKHWLF FDVH LV VXEMHFWHG WR UDQGRPO\ JHQHUDWHG SHUWXUEDWLRQV RI GLIIHUHQW PDJQLWXGHV 0DWWVVRQHWDO  7KHWRSRORJ\RIWKHHUURUIXQFWLRQZDVVWXGLHGIRUWKHV\QWKHWLFFDVH7KHVXUIDFHZDVIRXQG WREHQRQVPRRWKZLWKDORWRIQRLVHVDQGIOXFWXDWLRQVDQGDODUJHQXPEHURIORFDORSWLPDZHUH VSUHDGRYHUWKHHQWLUHVHDUFKGRPDLQ,WFDQEHFRQFOXGHGWKDWWKHJHQHWLFDOJRULWKPLVDSURSHU FKRLFH DV VHDUFK PHWKRG IRU WKLV URFNILOO GDP DSSOLFDWLRQ VLQFH WKH JHQHWLF DOJRULWKP LV QRW UHVWULFWHG E\ VPRRWKQHVV SURSHUWLHV RI WKH HUURU IXQFWLRQ VXUIDFH DQG LV QRW HDVLO\ WUDSSHG LQWR ORFDORSWLPD0DQ\RWKHUW\SHVRIVHDUFKPHWKRGVOLNHHJJUDGLHQWDQGGLUHFWVHDUFKPHWKRGV FDQEHH[SHFWHGWRKDYHSUREOHPVZLWKWKHHUURUIXQFWLRQWRSRORJ\LQWKLVDSSOLFDWLRQ Vol. 18 [2013], Bund. X 5437

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¶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

ACKNOWLEDGEMENTS 7KH UHVHDUFK SUHVHQWHG LQ WKLV SDSHU ZDV FDUULHG RXW DV D SDUW RI 6ZHGLVK +\GURSRZHU &HQWUH69&69&KDVEHHQHVWDEOLVKHGE\WKH6ZHGLVK(QHUJ\$JHQF\(OIRUVNDQG6YHQVND .UDIWQlW WRJHWKHU ZLWK /XOHn 8QLYHUVLW\ RI 7HFKQRORJ\ 7KH 5R\DO ,QVWLWXWH RI 7HFKQRORJ\ &KDOPHUV8QLYHUVLW\RI7HFKQRORJ\DQG8SSVDOD8QLYHUVLW\ 3DUWLFLSDWLQJK\GURSRZHUFRPSDQLHVDUH$QGULW]+\GUR,QHSDU6ZHGHQ$QGULW]:DSODQV (21 9DWWHQNUDIW 6YHULJH )RUWXP *HQHUDWLRQ  +ROPHQ (QHUJL -lPWNUDIW .DUOVWDGV (QHUJL /LQGH (QHUJL 0lODUHQHUJL 6NHOOHIWHn .UDIW 6ROOHIWHnIRUVHQV 6WDWNUDIW 6YHULJH 6WDWRLO /XEULFDQWV6ZHFR,QIUDVWUXFWXUH6ZHFR(QHUJXLGH6YH0LQ8PHn(QHUJL9DWWHQIDOO5HVHDUFK DQG'HYHORSPHQW9DWWHQIDOO9DWWHQNUDIW$%9*3RZHUDQG:63

 Vol. 18 [2013], Bund. X 5438

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p   %DFN DQDO\VLV RI JHRPHFKDQLFDO SDUDPHWHUV LQ XQGHUJURXQG ZRUNV XVLQJ DQ (YROXWLRQ 6WUDWHJ\ DOJRULWKP 7XQQHOOLQJ DQG 8QGHUJURXQG 6SDFH 7HFKQRORJ\YROSS(OVHYLHU  0XLU :RRG '   *HRWHFKQLFDO 0RGHOOLQJ 7D\ORU  )UDQFLV 6SRQ 3UHVV 2[IRUGVKLUH  1HOGHU-$DQG0HDG5  $VLPSOH[PHWKRGIRUIXQFWLRQPLQLPL]DWLRQ 7KH&RPSXWHU-RXUQDOYROQRSS  3DO 6 :LMH :DWKXJDOD * DQG .XQGX 6   &DOLEUDWLRQ RI D &RQVWLWXWLYH 0RGHO XVLQJ *HQHWLF $OJRULWKPV &RPSXWHUV DQG *HRWHFKQLFV YRO  QR  SS   3DSRQ $ 5LRX < 'DQR & DQG +LFKHU 3   6LQJOH DQG PXOWLREMHFWLYH JHQHWLFDOJRULWKPRSWLPL]DWLRQIRULGHQWLI\LQJVRLOSDUDPHWHUV,QWHUQDWLRQDO-RXUQDO IRU1XPHULFDODQG$QDO\WLFDO0HWKRGVLQ*HRPHFKDQLFVYROQRSS  3OD[LV   )LQLWH (OHPHQW &RGH IRU 6RLO DQG 5RFN $QDO\VLV YHUVLRQ  KWWSSOD[LVQO  3RWWV '0 DQG =GUDYNRYLü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ol. 18 [2013], Bund. X 5440

2SWLPL]DWLRQWK,&2/'(XURSHDQFOXEV\PSRVLXP6KDULQJ([SHULHQFHIRU6DIH DQG6XVWDLQDEOH:DWHU6WRUDJH$SULO9HQLFH,WDO\7RSLF%SS 

 ‹(-*( 

Paper C

Inverse Hardening soil parameter identification of an earth and rockfill dam by genetic algorithm optimization

Vahdati, P., Levasseur, S., Mattsson, H. and Knutsson, S

To be submitted to an international Journal

Inverse Hardening soil parameter identification of an earth and rockfill dam by genetic algorithm optimization

Pooya Vahdati

Department of Civil, Environmental and Natural resources engineering, Luleå University of Technology, Tel: +46 920 491764, SE – 971 87, Luleå, Sweden e-mail: [email protected] Séverine Levasseur Doctor Département ArGEnCo Service de Géomécanique et géologie de l'ingénieur Université de Liège, Tel: +33 249 5324067, 4000 Liège-Belgium, e-mail: [email protected]

Hans Mattsson Assistant Professor

Department of Civil, Environmental and Natural resources engineering, Luleå University of Technology, Tel: +46 920 492147, SE – 971 87, Luleå, Sweden e-mail: [email protected]

Sven Knutsson Professor

Department of Civil, Environmental and Natural resources engineering, Luleå University of Technology, Tel: +46 920 491332, SE – 971 87, Luleå, Sweden e-mail: [email protected]

Abstract

This paper presents a study of identification of constitutive parameter values in the Hardening soil model by inverse analysis of an earth and rockfill dam application. The authors have experience from a previous study on the same case with the Mohr-Coulomb model. The objective of this research is to study if the inverse analysis technique can be successfully used for this type of application and choice of constitutive model. The values of soil parameters are determined based on horizontal deformations obtained from installed instrumentations in the dam. The quantities that are monitored in the dam can be numerically predicted by a finite element simulation. In inverse analysis, constitutive parameter values are chosen in such a way that the error between data recorded by measurements in the dam and numerical simulation is minimized. An optimization method based on the genetic algorithm was applied to search for the minimum error in the search domain in this study. Optimizations have initially been performed in a large search domain in order to find a criterion identifying the best solutions. Thereafter, the optimizations were limited to this criterion in order to find the best set of solutions close to the

1 optimum point. Moreover, the error function topology and smoothness was examined as well. It was overall concluded, that the inverse analysis technique could be effectively used for earth and rockfill dam applications, despite the fact that the technique is expensive in terms of computational time.

KEYWORDS: Earth and rockfill dam, finite element modelling, inverse analysis, genetic algorithm, Hardening soil model.

Introduction

Computational modelling of earth and rockfill dams might not be an easy task if limited information is available about the material properties in the dam zones, due to lack of reliable relevant data which is the case for many old dams. In addition, taking samples for testing from dam zones, especially from the core, the central impervious part, is normally very difficult since it might affect the dam performance and the safety.

Inverse analysis can provide a method to determine the constitutive behaviour of various materials within the dam structure if the dam is well equipped with different instrumentations. In the method of inverse analysis, two separate parts are included: (1) an optimization method consisting of an error function and a search algorithm and (2) a numerical method to solve the partial differential equations arising in stress-strain analysis of structures. In this study, the numerical modelling was performed with a commercial finite element program Plaxis (2011) and the genetic algorithm was utilized as the search algorithm in the optimization method. The genetic algorithm was chosen due to its robustness, efficiency and particularly since it provides a set of solutions close to the optimum solution instead of one unique answer; a set of solutions is more practical from a geotechnical perspective. With an inverse modelling approach, a finite element model is calibrated step by step by changing the input values of the constitutive model or models until the difference between the measured data by the dam instrumentations and the associated computed values is minimized. The inverse analysis technique was introduced to the geotechnical field by Gioda and Sakurai (1987) for the purpose of identification of elastic material properties of in situ rock masses. Inverse analysis has been widely used recently thanks to the availability of sufficiently fast computer hardware and useful software. In the geotechnical engineering field the interest of applying inverse parameter identification strategies and optimization algorithms modelling in order to make inverse analysis procedures automated is growing fast (Swoboda et al. 1999, Gioda et al. 1980, Tarantola 2005, Calvello et al. 2002 and 2004, Finno et al. 2005, Ledesma et al. 1996, Lecampion et al. 2005, Lecampion et al. 2002, Zentar et al. 2001, Yu et al. 2007, Moreira et al. 2013, Levasseur et al. 2008, 2009 and 2010, Rechea et al. 2008, Papon et al. 2012).

Soil behavior is very complicated and can vary considerably under different loading conditions. Fundamentally, soil behaves as an elasto-plastic material. No constitutive model can completely describe the complex behavior of different soils in all loading situations. Therefore, constitutive

2 models have limitations to their applicability. All models have certain advantages and limitations which highly depend on their particular application. Accurate modelling of deformation and strength in earth and rockfill dams requires a complex stress-strain model. Duncan (1996) recommends the use of models that incorporate plasticity theory for modelling undrained behavior, e.g. for wet placed earthfill cores of thin to medium width in zoned embankments. From the deformation modelling of a concrete face rockfill dam (CFRD), Kovacevic (1994) draws the conclusion that elasto-plastic constitutive models are more suited to accounting realistically for the rockfill deformation behavior under the stress path conditions imposed. In this research, the constitutive relations of the zones of soil were represented by the Hardening soil model. This model was originally developed from the Duncan-Chang hyperbolic model (Duncan et al. 1970, Duncan 1981). However, this model, by using plasticity theory instead of elasticity theory could overcome the restrictions of hyperbolic models. The Hardening soil model includes dilatant soil behavior and introduces a yield cap. Since in reality the stiffness of soils depends on the stress level, the Hardening soil model is formulated to capture this stress level dependency.

The stress and strain behavior in embankment dams, which is governed by loads, particularly: dead weight of soil layers, water reservoir pressure and seismic activity, is predicted from the strength variation and the deformation behavior of soil under combined stresses. Deformations in embankment dams are due to changes in effective stress during construction, impoundment and reservoir fluctuations. The high horizontal stress from the water load acting on the upstream face of the core results in greater changes in lateral than vertical stress in the core and downstream shoulder. Therefore, the deformations will be significantly in the horizontal direction. The magnitude of the deformations depends on the changes in stress conditions as well as the compressibility and stiffness properties of the core and downstream rockfill. Based on studies of several monitored dams, Hunter (2003) shows that dams with poorly compacted rockfill in the downstream shoulder or embankments with cores of silty sands to silty gravels had great lateral crest displacements due to the reservoir impoundment.

The objective of this paper is to study if the inverse analysis technique can be effectively used for identification of constitutive parameter values in the advanced Hardening soil model, in an earth and rockfill dam application. The robustness and efficiency of the inverse analysis procedure and the genetic algorithm as a search algorithm were examined in a previous study, performed by the authors, based on the Mohr-Coulomb model for a synthetic case, Vahdati et.al (2013). Therefore, in this research it was focused only on the real dam case with the Hardening soil model as the constitutive model. In this study, geometry and zoning as well as inclinometer data from an existing earth and rockfill dam were used, see figure (1).

In this research, the inverse analysis was performed against two situations named case A and B, regarding the reservoir water level and the time for the construction of the berms. Moreover, the topology of the error function surface was studied to obtain knowledge about its shape and how the individual solutions in the optimized set of solutions become distributed on the function surface for the two cases. At the end, a proper range of values for the optimized parameters was

3 defined based on inclinometer data and the associated numerically simulated values. Notice that the optimizations presented in this paper are voluntarily restricted to two optimization variables.

A very good agreement from the best set of solutions was obtained against the inclinometer measurements for both cases A and case B. The Hardening soil model captured the pattern of horizontal displacements quite well.

The earth and rockfill dam

The earth and rockfill dam used for the case study is 45m high, with a dam crest of 6.5m, see figure (1). The downstream slope is inclining 1V:1.85H and the upstream slope is inclining 1V:1.71H. The impervious core, consisting of moraine materials, is surrounded by fine and coarse filters and a supporting fill of blasted rock. The dam was strengthened in two stages with rockfill berms on the downstream part. The first stabilizing berm of rockfill was constructed in 1990 and the second berm was constructed three years later in 1993. The dam is well equipped with instrumentations and monitoring programs including deformation measurements within the dam structure. Horizontal displacements are measured by an inclinometer installed along the downstream fine filter, from the crest into the rock foundation. Since the inclinometer data from the ten first meters from the bottom of the core seem not to be trustworthy, only data above this level are considered. The level of water in the reservoir annually changed from +430m to +440m.

Figure 1: Cross section of the dam

Finite element model

The Plaxis finite element program (2011) was applied to the numerical analyses in this study. The finite element model of the cross section of the dam is shown in figure (2). Because the dam is a long structure, a plane strain model was presumed. Some smaller simplifications compared to the real dam cross section in figure (1) were made in the finite element model in figure (2), e.g. the upstream toe berm was neglected due to similarity in its material properties with rockfill. Horizontal filters were ignored as well, since consolidation time is not taken into account. These simplifications had no practical influence on the modelling of this dam. For the geometry boundary conditions which are provided by fixities in Plaxis, the horizontal displacement of the outermost vertical left and right foundation boundary as well as the horizontal and vertical

4 displacement of the lowest horizontal boundary was prescribed to zero. The finite element mesh type was chosen as 15-noded triangular elements with twelve-Gauss points for the integration. Computation time is an important issue in inverse analysis, so the coarseness of the mesh was selected so that a sufficient accuracy of the computations was obtained for a minimum of computation time spent. Several different meshes were tested and the one showed in figure (2) was chosen. The dam embankment in the model in figure (2) was built up stepwise in five stages in order to generate a proper initial effective stress field (Duncan 1996, Reséndiz et al. 1972, Clough et al. 1967). The excess pore pressures in the core were dissipated by consolidation after construction of the dam. For this study two cases (A and B) based on different reservoir water levels and the number of berms constructed, were analyzed. Case A, refers to the time when the first stabilizing berm was constructed in 1990 and the reservoir water was at the minimum level of +430m. Case B, refers to when the reservoir water level reached its maximum level +440m and the second berm was just finished. The second berm was rapidly constructed right after the water level reached +440m. All the data of horizontal displacements from finite element computations received from exactly the same positions in the geometry as the measurements carried out with the single inclinometer in figure (1) were compared to the related inclinometer measurements in the expression for the error function. In the finite element analysis, all displacements were reset to zero before impoundment.

Case A Case B

Figure 2: Finite element model of the dam for case A and case B

Constitutive law; Hardening soil model

The Hardening soil model is a cone-cap model (Schanz et al. 1999). The plastic strain behavior on the cone is described by means of a shear yield surface following a non-associated flow rule based on Rowe’s stress-dilatancy theory (1962). The plastic potential is defined to assure a hyperbolic stress/strain response for triaxial compression loading. The cap yield surface is controlled by the volumetric plastic strain. On this cap, an associated flow rule is assumed. This yield surface is defined in order to close the elastic region in the -axis direction. The failure on the cone part is defined based on Mohr-Coulomb’s failure criterion′ which is characterized by the 푝

5 cohesion c', the friction angle ϕ' and the dilatancy angle ψ'. The Hardening soil model use a hyperbolic stress-strain curve on the cone instead of a bi-linear relation as in the Mohr-Coulomb model.

The Hardening soil model has stress-dependent moduli parameters, E50 , Eoed and Eur, which are defined by a power law parameter m. E50 , Eoed and Eur are the secant stiffness in primary triaxial loading, tangent (oedometer) stiffness and unloading/reloading stiffness, respectively. These ref ref ref stress dependent moduli E50 , Eoed and Eur are introduced in Plaxis as parameters associated to a reference pressure ( =100 kPa). The power law parameter m controls the shape of the cone - cap yield loci and m i푟푒푓s varied from 0.5 for a typical hard soil to 1 for a clay soil. The parameter 푃 values of γ u γ s , ϕ', kx and ky (quantities are explained in table 2) for all soil layers in the dam were provided by the dam owner. Proper values of c' were chosen based on advice in Bowles (1988). All the dilatancy angles were calculated from the same empirical relation, commonly accepted for dense sand: ψ' = ϕ'-30° as proposed by Plaxis (2011). The unloading/reloading Poisson’s ratio , is usually set between 0.15-0.30. Table (1) presents the conventional way of estimating parameters of the Hardening soil model as well as some appropriate default values based on 휈푢푟 recommendations from Plaxis (2011).

Schanz and Vermeer (1998) showed that based on standard drained triaxial tests the reference ref secant stiffness E50 of sand varied from 15 MPa for loose sand to 50 MPa for dense sand. In this ref study, the filter materials are close to dense sand, therefore the value of E50 was chosen to 50 ref MPa for these zones. The range of values for the reference secant stiffness of the core E50C and ref the reference secant stiffness of the rockfill material E50R for both case A and case B, were obtained from optimizations against the dam application in this study. A high value of the reference secant stiffness was considered for the foundation to make it rigid. Table (2) summarizes the values of the material properties of the rockfill dam which are utilized in this case study.

The dam core was modeled with the option of undrained case A in the Plaxis software, an option in which effective strength parameters are utilized. This mode was chosen in order to capture the hardening behavior of the soil. Undrained case B (undrained strength, ϕ' = 0 and c' = ) as utilized by the authors Vahdati et.al (2013) in a previous study, is not able to capture the stress 푐푢 dependency behavior in the Hardening soil model and the response would become as in the Mohr-Coulomb model (constant stiffness). All the other dam zones were modeled with the option of drained behavior, i.e. no excess pore pressures are generated.

6

Table 1: Parameters of the Hardening soil model

Parameter Definition Relation

ϕ' Friction angle Slope of the failure line in τ - stress plane ′ 휎 * Failure ratio = 푞푓 푅푓 푅푓 푞푎 ψ Dilatancy angle ψ = ϕ' -30° ′ ′ ' ' m ref Reference secant stiffness for primary ref c cot ϕ'-σ3 *E E50 = E 50 loading in a drained triaxial test 50 c' cot ϕ'-Pref � � ' ' m ref Reference tangent stiffness for primary ref c cot ϕ'-σ1 *E Eoed = E oed oedometer loading oed c' cot ϕ'-Pref � � Reference stiffness for m c' cot ϕ'-σ' *Eref unloading\reloading in a drained triaxial E = Eref 3 ur ur ur ' ref test c cot ϕ'-P � � Power for stress-level dependency of σ' m Slope of trend line in log 3 - log E plane stiffness Pref 50 � � *νur Poisson's ratio for loading\unloading 0.15-0.30

' σ3 K0 Earth pressure coefficient at rest K0= ' σ1

Earth pressure coefficient value for Knc Knc= 1-sin 0 normally consolidated soil 0 휙′ Note: is the ultimate deviatoric stress, is the asymptotic deviatoric stress

푞푓 ref ref ref ref 푞푎 * E50 =Eoed, Eur = 3E50 , = 0.9 and νur= 0.2 in this study

푅푓

7

Table 2: Material properties of the rockfill dam

ref ref ref γu γs E50 =Eoed Eur m c' ϕ' kx/ky Dam zones kN/m3 MPa MPa - kPa (°) m/s

Core 21 23 70-100 210-300 1 20 38 3.0·10-7

Fine filter 21 23 50 150 0.5 0 32 9.0·10-5

Coarse 21 23 50 150 0.5 0 34 5.0·10-4 filter

Rockfill 19 21 10-17 30-51 0.5 7 30 1.0·10-2

Berms 21 23 10 30 0.5 7 30 5.0·10-2

Foundation 21 23 3000 9000 0.5 0 45 1.0·10-8

Note: γ u is the unit weight above the phreatic level,γ s is the unit weight below the phreatic ref ref level, E50 is the Reference secant stiffness for primary loading in a drained triaxial test, Eoed is ref the Reference tangent stiffness for primary oedometer loading, Eur is the Reference stiffness for unloading\reloading in a drained triaxial test, m is the Power for stress-level dependency of stiffness, c' is the effective cohesion, ϕ' is the effective friction angle and kx and ky are the hydraulic conductivity in horizontal and vertical direction, respectively.

Optimization method

The aim of optimization is to reach the optimum solution among other solutions for a given problem in a matter of efficiency. For any kind of optimization an error function and a search method are needed. In this paper, the error function provides a scalar measure of the discrepancy between horizontal deformations obtained by inclinometer measurements and numerical simulations, respectively. The genetic algorithm was chosen as the search method for the purpose of finding the minimum value of the error function. The variables for optimizing the selected parameters are included in constitutive models utilized for the finite element computations. Notice that, the magnitude of the error function is thereby influenced by the values of the optimization variables. The optimal values of the chosen parameters are progressively approached through iteration by minimizing the error function; that is the optimum solution.

8

Error function

In order to define the difference between the measured values and finite element computed values, it is necessary to define a proper error function. In this research, following Levasseur et

al. (2008) a scalar error function Ferr based on the least-square method was introduced

1 N 2 2 1 Ue - Un i i � (1) Ferr = ΔU 2 i=1 � i � � � � 푁 that determines the difference between experimentally measured values Ue i and numerically

computed values Uni , from the N measuring points. The term 1/ΔUi gives the weight of the

difference betweenUe i and Uni. The parameter ΔUi relates to the numerical and experimental uncertainties at the measurement point i; defined as

Ui = ε + αUe i (2)

where the parameter ε is an absolute훥 error of measurements and the parameter α is a dimensionless relative error of measurements.

In this study the error function defined in equation (1) is considered with the parameters ε = 1 and α = 0. That is to say, the function, given by equation (1) is reduced to

1 N 2 1 2 F = U - U � (3) err N e i ni i=1 � � � � � and has the unit of length.

Search algorithm

The genetic algorithms were developed originally in the field of artificial intelligence by John Holland in 1970s at the University of Michigan and popularized as a universal optimization algorithm; the method was inspired by Darwin’s theory of evolution. The genetic algorithms are computationally and numerically simple and powerful in their search for improvement. The genetic algorithm is a global optimization technique which is based on the genetic mechanisms, such as reproductions, crossings and mutations, with the aim of localizing an optimum set of solutions close to the optimum without being trapped into local optima in the given search domain; but without any guarantee of finding the exact global optimum (Pal et al. 1996, Gallagher et al. 1991, Haupt 1998). Furthermore, this search algorithm is not fundamentally restricted by smoothness properties of the error function surface, like continuity and existence of

9 derivatives. The optimization program that is used in this research was developed in the FORTRAN language by Levasseur et al. (2008, 2009 and 2010) for geotechnical studies.

In the genetic algorithm, the space of search is defined based on the number of parameters Np to optimize. The minimization problem is approached in the Np dimensional space limited by the upper and the lower bound constraints of each parameter. The value of each optimization parameter is binary encoded to a form of gene. Each gene is encoded into a part of the bit string in order to be passed to the numerical model. The combination of the genes forms individuals (or chromosomes). An individual is a genetic algorithm chromosome which has raw genetic information Individuals are the single solutions. The population’s size Ni consists of the set of individuals, solutions, involved in the search domain. First, an initial population, twice the population size in this study, is defined by random choice of values of chromosomes in Np domain. The initial population represents the start of the search for the optimum, and explores in the search space spanned by the optimization variables and their boundaries. Initially, these random solutions need to be evaluated. Based on the individual’s fitness value, the population is sorted in ascending or descending order. In the genetic algorithm the fitness is the value of the error function for its individual. Only Ni /3 of the best individuals are selected for the next population, which are called parents. Randomly selected pair parents, good temporary solutions, produce a new generation (offspring), which get some parts of the parents’ genetic material, based on crossover or mating. The most common form of mating includes two parents that produce two children (offspring). The position of crossover points is randomly chosen and the value of each binary string exchanges with another one after this position point. Each parent copies e.g. half the gene to each child, with the selection of the gene part being chosen randomly. In order to improve the efficiency of the algorithm a crossover point number is chosen equal to the number of selected parameters, which was proposed by Pal et al. (1996). This process is repeated until 2Ni/3 children are created. Now the new population, consisting of parents and children is completed. The procedure of selection, crossing and reproduction is called breeding. Mutations apply for preventing the genetic algorithm from converging too fast before sweeping the entire search domain. Mutation is the variation of a randomly selected bit in the parameter code based on a specific probability.

Population size

A proper population size must be chosen before running any genetic algorithm computation. The size of the population must initially be large enough to be effectively spread out over the search space. To avoid a prohibitive CPU time, it is then necessary to dynamically reduce the size of this population (Chelouah et al. 2000). Also, large population sizes would be required to reduce the risk of solutions getting trapped in local optima on the noisy and fluctuated error function surface. However, the drawback of large population sizes is the computation time cost which in some cases could be equivalent to a random search. A small population would be more likely to quickly converge, which might be to a local optima, and the quality of convergence is left to

10 chance. A population sizing equation from statistical theory proposed by Goldberg (1991) and analyzed by Goldberg (1991) as well as Carroll (1996) was studied in this research

npop = = (4) 푘 푘 where X is the number of possibilities for each chromosome,푙 e.g. for binary X=2, is the length of 푚 푋 �푘 푋 chromosome and k is the size of schema, details about the schema can be found in Goldberg 푙 (1989).

To estimate a proper population size, it was considered that each parameter string would represent one schema. So, the length of the schema was assumed to be equal to parameter length for this purpose. For this study, =12, X=2 and k=6. Therefore the proper population size based on this theory would be 푙 12 6 npop= 6 2 =128 (5) The population size which was chosen �based� � on� numerical� tests in Vahdati et.al (2013), agrees well with results obtained with equation (5).

Inverse analysis of case A and case B

This study is a continuation of a previous study, Vahdati et.al (2013), for the purpose of improving the choice of constitutive model from the simple Mohr-Coulomb model to the advanced Hardening soil model to monitor whether it is possible to approach a better agreement with inclinometer data of horizontal displacements. The finite element models of cases A and B are presented in figure (2). In the previous study, only case B was considered with the incremental stiffness theory of the Mohr-Coulomb model. According to available data and results from that study, the smallest average error function value is considered as an upper limit of the error function value for this study, in order to monitor if the advanced Hardening soil model is able to approach better results compared to the Mohr-Coulomb model. For case A, no former studies were made and an upper limit for the error function value needs to be defined. Therefore, before starting to utilize the Hardening soil model, as the main constitutive model in this research, the upper limit of the error function value for case A should be defined.

Mohr-Coulomb model and case A

core rockfill In the previous study, Ginc and Ginc were the two optimization variables for the inverse analysis based on the Mohr-Coulomb model. Moreover, the population size of 120 was selected as the best population size for the inverse analyses which is also suggested theoretically above. Therefore, based on these assumptions, for the purpose of defining the lowest error function value for case A, initially, ten inverse analyses were performed in a search space of core rockfill [Ginc ,Ginc ] restricted to [10 kPa - 5000 kPa] , in order to get an idea about the location of probable solutions. Then, after the proper range of potential solutions was defined, the boundary was further limited to the range [2500 kPa - 4660 kPa] for the core and [10 kPa – 280 kPa] for the 11 rockfill. Figure (3) presents the best solution set obtained. It can be concluded that the lowest -3 error function value is Ferr = 1.50·10 which can be considered as the upper error function limit for case A.

Figure 3: Best solution set for case A with the Mohr-Coulomb model

Hardening soil model

Before optimization, a sensitivity analysis of the model parameters in table (1) was utilized to characterize the influence of each parameter. From this, it was found that the reference secant ref ref stiffness of the core E50C and the reference secant stiffness of the rockfill E50R were two of the most sensitive parameters for the horizontal deformation in the downstream zone of the dam. Therefore, these two parameters were chosen as the optimization variables in this study.

Ten inverse analyses were performed with the population size of 120 for the two cases A and B. In order to have a broad perspective about the location of the probable solution sets within the ref ref ref ref domain of [E50C, E50R ], large limits [5 MPa-100 MPa] were chosen for E50C and E50R. The best -3 error function value Ferr = 1.7·10 obtained from the previous study by Vahdati et.al (2013), and is considered as an upper limit for the acceptable solutions for case B. For case A, as previously -3 defined, an error function value of Ferr = 1.5·10 is considered as an upper limit for acceptable solutions. All the solutions which have an error function value less than the defined upper limit for each case, have the potential of being answers. In this research, each optimization parameter

12 is encoded by 6 bytes in the genetic algorithm. Then the search domain consists of 26 times 26 nodes, 4906 discrete points with a mesh resolution of Δ1.48 MPa.

Figures (4) and (5), present the set of solutions below the upper limit of Ferr, of ten inverse analyses for the cases A and B, respectively. One can conclude that for case A, the acceptable ref solutions for the value of E50R is limited to [13 MPa-18 MPa], while this range is from [55 MPa- ref ref 95 MPa] for E50C . On the other hand, for case B this range for E50R is limited to [10 MPa-20 ref MPa] and for E50C to [20 MPa-100 MPa]. Moreover, from these analyses it can be concluded that ref ref the horizontal deformation is more sensitive to the value of E50R than E50C , since the range of ref ref acceptable solutions became more limited for E50R than E50C . As can be seen in figures (4) and (5), the solutions differ with an integer multiple of 1.48 MPa, which is related to the mesh size chosen for the search domain.

Set of solutions of ten inverse analyses for case A

95

90

85

80

75

70

65 Reference secant stiffness corethe [MPa]of stiffness secant Reference 60

55 13 14 15 16 17 18 Reference secant stiffness of the rockfill [MPa]

Figure 4: Set of solutions of ten inverse analyses for case A

13

Set of solutions of ten inverse analyses for case B 100

90

80

70

60

50

40 Reference secant stiffness corethe [MPa]of stiffness secant Reference 30

20 10 11 12 13 14 15 16 17 18 19 20 Reference secant stiffness of the rockfill [MPa]

Figure 5: Set of solutions of ten inverse analyses for case B

Figure (6) presents the evaluation of the average error function value Ferr versus the number of finite element computations for one chosen inverse analysis of cases A and B, � respectively. This results show how the average error function value decrease when the set of solutions is gradually improved in the process of optimization. The process of optimization begins from a random initial population (240 computations) and continues until it converges to the best set of solutions. It can be seen, that for case A, the process converges after seven generations and 424 finite element computations to the value of -3 Ferr = 1.40·10 and for case B the process converges after six generations and 392 finite element computations to F = 1.29·10-2. Convergence is reached when the optimization � err process is no longer able to approach better results, then the average error function value � gets more or less constant for each new generation of a new population, marked with square symbols in figure (6). The number of finite element computations is closely related to computation time.

14

0.0050 0.0200

0.0045 0.0190

0.0040 0.0180

0.0035 0.0170

0.0030 0.0160

0.0025 0.0150

0.0020 0.0140 Average error function value Average error function value 0.0015 0.0130

0.0010 0.0120 240 280 320 360 400 440 480 240 260 280 300 320 340 360 380 400 Number of finite element computations Number of finite element computations case A case B

Figure 6: Average error function values versus number of finite element computations for the large search domain

Study of error function topology

The purpose of this study is to find a graphical shape of the error function in order to monitor the smoothness properties of the surface and examine the location of the global optimum as well as the distribution of local optima in the search domain. Since only two optimization variables are involved, it is convenient to study the topology of the error function graphically in this case.

Figures (7) and (8) represent the topology and the level contours of the error function for case A and case B, respectively. The error function values Ferr were plotted in the search domain limited ref ref to [5 MPa-100 MPa] for the optimization variables E50C and E50R. For both cases A and B, a ref distinct valley is visible on the error function surface in the low range of E50R values [10 MPa-20 MPa]. This valley that contained the location of the global optimum is extended with a slight ref inclination along the whole search range for E50C values [5 MPa-100 MPa]. The error function surface of case A has smaller fluctuations and noises with a number of local optima spread over the entire search domain, while the error function surface of case B is smooth. Therefore, the genetic algorithm is an appropriate choice for this application since the algorithm is not limited by smoothness properties of the error function and is not easily trapped into local optima.

15

(a) (b) Figure 7: case A (a) error function topology (b) level contours of error function

(a) (b) Figure 8: case B (a) error function topology (b) level contours of error function

Inverse analysis with reduced search domain of case A and B

ref ref In order to improve the optimization search within the [E50R , E50C ] space, the search domain is further limited to the range of the solution sets in figures (4) and (5). For case A, the search ref ref domain is limited to [13 MPa-18 MPa] for E50R and [55 MPa-95 MPa] for E50C with the mesh ref ref size of ΔE50R = 78.125 kPa and ΔE50C = 625 kPa, respectively. For case B, the search domain is ref ref limited to [10 MPa-20 MPa] for E50R and [20 MPa-100 MPa] for E50C with the mesh size of ref ref ΔE50R = 156.250 kPa and ΔE50C =1250 kPa, respectively. Since the mesh size of case B is twice the mesh size of case A, it is easier to compare the set of solutions from case A and case B, with each other.

16

Another ten inverse analyses, limited to the reduced domain, were performed in order to find the

best set of solutions for case A and B. Solutions are as defined before as points with Ferr < -3 -3 1.5·10 for case A and Ferr < 1.7·10 for case B and convergence is reached when Ferr gets more or less constant in the optimization process. � Figure (9) presents the topologies of the error function values of all solution points of the ten ref ref inverse analyses for both case A and case B within the reduced domain of [ E50C, E50R ]. For case A, there are two valleys in the defined domain area. The first one is located in the area of [75 ref ref MPa-100 MPa] for the E50C and [13.5 MPa-14.5 MPa] for the E50R and the second valley is ref ref located in the area of [70 MPa-100 MPa] for the E50C and [15.5 MPa-17 MPa] for the E50R . For case B, the best potential solutions are roughly located within the range of [20 MPa-100 MPa] for ref ref the E50C and [10.5 MPa-13 MPa] for the E50R.

The blue rectangle as well as the blue parallelogram is defined as the areas of the best solutions of case A and the red rectangle is defined as the best solutions of case B. It can be seen that in case B the value of the rockfill stiffness is lower compared to case A. It is difficult to evaluate the best solution, representative of both case A and case B at the same time based on figure (10), but ref it feels like such a solution should be taken somewhere in the range [70 MPa- 100 MPa] for E50C ref and [10.5 MPa-17 MPa] for E50R. It can be concluded that for being in both best solution’s ref domain, it is more probable that the E50C has a constant value during the time of happening these ref two cases than constant value for the E50R, see figure (10). Therefore, the most probable solutions ref are located in the constant value E50C within the range of [70 MPa- 100 MPa] for both cases with ref various values of the E50R within the range of [10 MPa- 17 MPa]. This figure also shows that in the case B the rockfill stiffness is lower compared to that in case A. However, in case A, the best ref ref solution gives the parameter values of E50C = 90 MPa and E50R = 14 MPa, with the error function ref ref value Ferr = 0.0011. For case B, the best solution becomes E50C = 90 MPa and E50R = 11 MPa with the error function value Ferr = 0.0125.

Figure (10) shows, for the best solutions for case A and B respectively, the measured inclinometer values and the associated numerically simulated values of the horizontal displacement at different heights of the dam. The good agreement for both cases demonstrates that the Hardening soil model is able to simulate the trends of horizontal displacements. Moreover, the progress of the average error function value versus the number of finite element computations for cases A and B were studied as well, see figure (12). For Case A, the average error function value decreased step by step in the optimization and after 457 finite element -3 computations and seven generations it converged to the value Ferr = 1.16·10 . For case B, after 284 finite element computations and five generations the average error function approached the -2 value Ferr = 1.25·10 .

17

ref ref Figure 9: Best set of solutions in the domain of [E50C, E50R ] for cases A and B 18

Ten set of solutions in the united domain of case A and case B Case B Case A 100

90

80

70

60

50

40

Reference secant stiffness corethe [MPa]of stiffness secant Reference 30

20 10 11 12 13 14 15 16 17 18 19 20 Reference secant stiffness of the rockfill [MPa]

ref ref Figure 10: Set of solutions in the united domain [E50C, E50R ] of cases A and B

Case A- inclinometer values FE-model results of case A values Case B- inclinometer values FE- model results of case B values 45

40 ] 35

30

25

20

15 The height ofthe dam [m 10

5 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 Horizontal displacement [m] Figure 11: Measured and computed horizontal displacements at different heights of the dam for cases A and B

19

0.00125 0.01290

0.00124 0.01285 0.00123 0.01280 0.00122 0.01275 0.00121 0.00120 0.01270

0.00119 0.01265 0.00118 0.01260

0.00117 Average error function value Average error function value 0.01255 0.00116 0.00115 0.01250 240 280 320 360 400 440 480 240 250 260 270 280 290 300 Number of finite element computations Number of finite element computations Case A Case B Figure 12: Average error function values versus number of finite element computations for the reduced search domain

Concluding remarks

This research was focused on validating the technique of inverse analysis with the aim of identifying parameter values of constitutive models in an earth and rockfill dam application. The constitutive model chosen in this study was the Hardening soil model. The genetic algorithm was applied as the search method in the optimization.

Two case studies named A and B, regarding different reservoir water levels and numbers of berms constructed were considered for inverse analyses. Ten inverse analyses were performed for each case with the population size of 120 which was chosen based on a previous study by Vahdati et.al (2013). Initially, in order to have a broad perspective about the location of the probable solution set within the domain, large limits were chosen for the optimization variables. From these initial optimizations, a reduced search area was defined for cases A and case B.

Thereafter, another ten inverse analyses were applied to each case in the reduced domain. Smaller mesh size in the search domain of each case of the set of inverse analyses led to the approach of more solution points. Therefore, the probability of finding better solutions is increased. The progress of the average error function value versus the number of finite element computations was studied for the ten inverse analyses in each case. It was seen that by reducing the search domain, better sets of solutions with lower error function values could be approached. A very good agreement for the simulation with the best solution was obtained against the inclinometer measurements for both case A and case B. The Hardening soil model captured the trend of horizontal displacements quite well.

20

The topology of the objective function for case A was found to be fluctuated while the surface for case B was found to be smooth without noticeable fluctuations. From this study and the previous study it can be concluded that the smoothness of the error function shape not only depends on the constitutive model chosen, smoothness depends also on the optimization case and the resolution in the search domain. Moreover, it can be concluded that the genetic algorithm is an appropriate option as search method for this rockfill dam application, since the genetic algorithm is not limited by smoothness properties of the error function surface and is not easily trapped into local optima. Many other types of search methods, like e.g. gradient and direct search methods, can be expected to face problems with the error function topology in this application.

In this research, two optimization variables were considered. An increased number of optimization variables will be considered for future work.

Acknowledgements The research presented in this paper was carried out as a part of "Swedish Hydropower Centre - SVC". SVC has been established by the Swedish Energy Agency, Elforsk and Svenska Kraftnät together with Luleå University of Technology, The Royal Institute of Technology, Chalmers University of Technology and Uppsala University.

Participating hydro power companies are: Andritz Hydro Inepar Sweden, Andritz Waplans, E.ON Vattenkraft Sverige, Fortum Generation, Holmen Energi, Jämtkraft, Karlstads Energi, Linde Energi, Mälarenergi, Skellefteå Kraft, Sollefteåforsens, Statkraft Sverige, Statoil Lubricants, Sweco Infrastructure, Sweco Energuide, SveMin, Umeå Energi, Vattenfall Research and Development, Vattenfall Vattenkraft AB, VG Power and WSP.

References 1. Bowles, J.E. (1988). "Foundation Analysis and Design". MacGraw-Hill. Berkshire. England. 2. Calvello, M. and Finno, R., (2002). "Calibration of soil models by inverse analysis". Numerical Models in Geomechanics NUMOG VIII, pp. 107-116. 3. Calvello, M. and Finno, R.J., (2004). "Selecting Parameters to Optimize in Model Calibration by Inverse Analysis". Computers and Geotechnics, vol. 31, no. 5, pp. 410-424. 4. Carroll, D.L., (1996). "Genetic algorithms and optimizing chemical oxygen-iodine lasers". Developments in Theoretical and Applied Mechanics, vol. 18, no. 3, pp. 411-424. 5. Chelouah, R. and Siarry, P., (2000). "A continuous genetic algorithm designed for the global optimization of multimodal functions". Journal of Heuristics, vol. 6, no. 2, pp. 191-213. 6. Clough, R.W. and Woodward, R.J., (1967). "Analysis of Embankment Stresses and Deformations". Journal of Soil Mechanics & Foundations Div. Vol 93, pp. 529-549. 7. Duncan, J.M. and Chang, C.Y., (1970). "Nonlinear analysis of stress and strain in soils". Journal of the Soil Mechanics and Foundations Division, vol. 96, no. 5, pp. 1629-1653.

21

8. Duncan, J.M., (1981). "Hyperbolic Stress-Strain Relationships". Application of Plasticity and Generalized Stress-Strain in Geotechnical Engineering. 9. Duncan, J.M., (1996). "State of the Art: Limit Equilibrium and Finite-Element Analysis of Slopes". Journal of Geotechnical Engineering, vol. 122, no. 7, pp. 577-596. 10. Finno, R.J. and Calvello, M., (2005). "Supported Excavations: Observational Method and Inverse Modelling". Journal of Geotechnical and Geoenvironmental Engineering, vol. 131, no. 7, pp. 826- 836. 11. Gallagher, K., Sambridge, M.S. and Drijkoningen, G., (1991). "Genetic Algorithms-an Evolution from Monte-Carlo Methods for Strongly Non-Linear Geophysical Optimization Problems". Geophysical Research Letters, vol. 18, no. 12, pp. 2177-2180. 12. Gioda, G. and Maier, G., (1980). "Direct Search Solution of an Inverse Problem in Elastoplasticity: Identification of Cohesion, Friction Angle and in Situ Stress by Pressure Tunnel Tests". International Journal for Numerical Methods in Engineering, vol. 15, no. 12, pp. 1823- 1848. 13. Gioda, G. and Sakurai, S., (1987). "Back Analysis Procedures for the Interpretation of Field Measurements in Geomechanics". International Journal for Numerical and Analytical Methods in Geomechanics, vol. 11, no. 6, pp. 555-583. 14. Goldberg, D.E., (1989). "Genetic algorithms in search, optimization, and machine learning". Machine learning 3.2 (1988): 95-99 15. Goldberg, D.E., Deb, K. and Clark, J.H., (1991). "Genetic algorithms, noise, and the sizing of populations". Complex Systems, vol. 6, pp. 333-362. 16. Haupt, R.L. and Haupt, S.E., (2004). "Practical Genetic Algorithms". Wiley Online Library. New York. 17. Holland, J.H., (1975). "Adaptation in Natural and Artificial Systems". University of Michigan Press: MIT press: Ann Arbor. 18. Hunter, G. and Fell, R., 2003. The deformation behaviour of embankment dams. Uniciv report No. R-416. The University of New Wales, School of Civil and Environmental Engineering 19. Kovacevic, N., (1994). "Numerical Analyses of Rockfill Dams, Cut Slopes and Road Embankments". PhD Thesis, London University (Imperial College of Science and Medicine), Faculty of Engineering. 20. Lecampion, B. and Constantinescu, A., (2005). "Sensitivity Analysis for Parameter Identification in quasi‐static Poroelasticity". International Journal for Numerical and Analytical Methods in Geomechanics, vol. 29, no. 2, pp. 163-185. 21. Lecampion, B., Constantinescu, A. and Nguyen Minh, D., (2002). "Parameter Identification for Lined Tunnels in a Viscoplastic Medium". International Journal for Numerical and Analytical Methods in Geomechanics, vol. 26, no. 12, pp. 1191-1211. 22. Ledesma, A., Gens, A. and Alonso, E., (1996). "Estimation of Parameters in Geotechnical back analysis—I. Maximum Likelihood Approach". Computers and Geotechnics, vol. 18, no. 1, pp. 1- 27. 23. Levasseur, S., Malecot, Y., Boulon, M. and Flavigny, E., (2009). "Statistical Inverse Analysis Based on Genetic Algorithm and Principal Component Analysis: Method and Developments using Synthetic Data". International Journal for Numerical and Analytical Methods in Geomechanics, vol. 33, no. 12, pp. 1485-1511.

22

24. Levasseur, S., Malecot, Y., Boulon, M. and Flavigny, E., (2008). "Soil Parameter Identification using a Genetic Algorithm". International Journal for Numerical and Analytical Methods in Geomechanics, vol. 32, no. 2, pp. 189-213. 25. Levasseur, S., Malecot, Y., Boulon, M. and Flavigny, E., (2010). "Statistical Inverse Analysis Based on Genetic Algorithm and Principal Component Analysis: Applications to Excavation Problems and Pressuremeter Tests". International Journal for Numerical and Analytical Methods in Geomechanics, vol. 34, no. 5, pp. 471-491. 26. Moreira, N., Miranda, T., Pinheiro, M., Fernandes, P., Dias, D., Costa, L., Sena-Cruz, J. (2013). "Back analysis of geomechanical parameters in underground works using an Evolution Strategy algorithm". Tunnelling and Underground Space Technology. vol. 33, pp.143-158, Elsevier 27. Pal, S., Wije Wathugala, G. and Kundu, S., (1996). "Calibration of a Constitutive Model using Genetic Algorithms". Computers and Geotechnics, vol. 19, no. 4, pp. 325-348. 28. Papon, A., Riou, Y., Dano, C. and Hicher, P. (2012). "Single- and multi-objective genetic algorithm optimization for identifying soil parameters. International Journal for Numerical and Analytical Methods in Geomechanics, vol. 36, no. 5, pp. 597-618. 29. Plaxis. (2011). Finite Element Code for Soil and Rock Analysis. version 2011-2.http://plaxis.nl 30. Rechea, C., Levassur, S. and Finno, R., (2008). "Inverse Analysis Techniques for Parameter Identification in Simulation of Excavation Support Systems". Computers and Geotechnics, vol. 35, no. 3, pp. 331-345. 31. Resendiz, D. and. Romo, M.P (1972). "Analysis of embankment deformations". Anonymous Performance of Earth and Earth-Supported Structures. ASCE. Vol.1. pp. 817-836 32. Rowe, P.W., 1962. "The stress-dilatancy relation for static equilibrium of an assembly of particles in contact". Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences, vol. 269, no. 1339, pp. 500-527. 33. Schanz, T. and Vermeer, P., (1998). "On the stiffness of sands". Géotechnique, vol. 48, pp. 383- 387. 34. Schanz, T., Vermeer, P., and Bonnier, P., (1999). "The hardening soil model: formulation and verification". Beyond 2000 in computational geotechnics: 10 years of PLAXIS International; proceedings of the International Symposium beyond 2000 in Computational Geotechnics, Amsterdam, The Netherlands, 18-20 March 1999. 35. Swoboda, G., Ichikawa, Y., Dong, Q. and Zaki, M., (1999). "Back Analysis of Large Geotechnical Models". International Journal for Numerical and Analytical Methods in Geomechanics, vol. 23, no. 13, pp. 1455-1472. 36. Tarantola, A., (1987). "Inverse Problem Theory". Elsevier. Amsterdam 37. Vahdati, P., Levasseur, S., Mattsson, H. and Knutsson, S., (2013)." Inverse Mohr-Coulomb soil parameter identification of an earth and rockfill dam by genetic algorithm optimization". Electronic journal of geotechnical engineering, vol. 18, pp.5419-5440. 38. Yu, Y., Zhang, B. and Yuan, H., (2007). "An Intelligent Displacement Back-Analysis Method for Earth-Rockfill Dams". Computers and Geotechnics, vol. 34, no. 6, pp. 423-434. 39. Zentar, R., Hicher, P. and Moulin, G., (2001). "Identification of Soil Parameters by Inverse Analysis". Computers and Geotechnics, vol. 28, no. 2, pp. 129-144.

23