Arxiv:1802.05745V1 [Physics.Chem-Ph] 15 Feb 2018 Rpitsbitdt Elsevier to Submitted Preprint Hoiso Ope Hmclratos He Simple Invented: Three Been Have Ideas Reactions

arXiv:1802.05745v1 [physics.chem-ph] 15 Feb 2018 rpitsbitdt Elsevier to submitted Preprint hoiso ope hmclratos he simple invented: Three been have ideas reactions. basic chemical first complex the of with theories together developed re- were model the methods that duction surprising not re- is model it Therefore, of technology duction. developed well impossi- the is without models working ble such of Creation [2]. works of modelling important. more of even became problem In reactions the complex engineering. context, chemical engineering of the needs the and ideas for on mathematical approaches of concentrated systematization was detailed activity the by in Aris’ Prize reactions” Nobel 1956). “chain shared Chemistry, of (the Hinshelwood concept since and the research Semenov dynamics of chemical invention of the focus the in wave.was Aris of Ho modelling the of vant problem the The (3) (2) the and 1901) Chemistry, (1) wave, in Semenov–Hinshelwood Prize Nobel [1]: first leaders (the wave their with associated introduction Historical Keywords: l tropic asymptotology, some result methods, new and bottleneck including art perturbation, kinetics, the biochemical of and state chemical current in the briefly describes It steps reduct (2) limiting for and methods manifolds, invariant and slow notions, equilibrium, ideas, basic presents It (1) Abstract • that device intellectual an is model mathematical A be can dynamics chemical of waves) (or eras Three goals: two has paper The mi address: Email oe euto nceia yais lwivratmanif invariant slow dynamics: chemical in reduction Model ateog n,atrta,rmi lotequili- almost remain that, equilibrium brated); after their and, enough approach fast reactions of fraction The etrain,temdnmcetmts n nlsso r of analysis and estimates, thermodynamic perturbations, Quasi-Equilibrium ecinntok us taysae quasi-equilibrium, state, steady quasi network, reaction [email protected] a eateto ahmtc,Uiest fLietr Leice Leicester, of University Mathematics, of Department ope ecinnetworks reaction complex prxmto rQ (a QE or approximation AN Gorban) (A.N. b oahvk nvriy ihiNvoo,Russia Novgorod, Nizhni University, Lobachevsky ..Gorban A.N. ff leulbain n ecinmcaimskeletonisation mechanism reaction and equilibration, al . nmtoso nain aiod,cmuainsingular computation manifolds, invariant of methods in s o frato ieismdl:qaised-tt,quas quasi-steady-state, models: kinetics reaction of ion ts civmnsi h ra rao oe reduction model of area broad the in achievements atest eea nesadn fmdlrdcini chemical in reduction model of understanding general [7]. anal- systems sensitivity kinetic the complex of and of approximation [6], ysis motions iterative fast and the slow for both method Perturba- (CSP) Singular tion Computational the of [5], approximation motion Dime- for slow Low method Intrinsic (ILDM) special Manifold the theory sional 4], (MIM) [3, Manifolds Invariant algorithms and of them Method of important the Most are: ideas. several by enriched was fol- the in presented sections. is lowing discussion formal precise More a,b • • eeomn flmigaayi a motn for important was analysis lumping of Development technology reduction model the 1990s – 1980s In ae ti igerato,wihrt sagood a is net- whole rate work. the of which rate reaction reaction, the to single approximation simplest a the is in it network, cases reaction the of part small iinta h ocnrtoso te pce do species other con- of change; under not concentrations state the steady that the QSS). dition as the defined to is close QSS of The remaining dynamics species the slave, other a these as follow, then concentra- and QSS of tions, their small quickly some relatively reach are in they exist amounts; these radicals, often or very intemediates species, of (some The The iiigse,dmnn ah nrp production entropy path, dominant step, limiting us tayState Steady Quasi iiigsteps limiting tr E R,UK 7RH, LE1 ster, rbtlnc sarelatively a is bottleneck or ato graph eaction ls singular olds, prxmto rQSS or approximation eray1,2018 19, February . i- kinetics [8]. The lumped species is considered as a lin- where ε> 0 is a small parameter. The fast subsystem is ear combination of the original ones. These combina- dx tions are often guessed on the basis of known kinetic = Ffast(x). (2) properties and can be improved by iterative methods dτ and observer theory. The standard examples are: (i) Here, time τ is used to stress that this is the ‘fast time’, the lumped species are identified as the sums of species τ = t/ε. If the fast system (1) converges to an asymp- in selected groups (a very popular approach with many totically stable fixed point in U and has no fixed points practical applications, for example, [9]); (ii) the lumped on the border of U then for sufficiently small ε the slow species are the numbers of links and structural frag- vector field becomes practically invisible in the dynam- ments of various types and in different states (this ap- ics of (1), i.e. there is no slow dynamics. proach has many applications, from petrochemitry [10] Let the fast system (2) be neither globally stable nor and modelling of intracellular networks [11] to the In- ergodic in U. Assume that (i) it has the conservation ternet dynamics [12]). laws bi(x)(i = 1,..., k) and (ii) for each x0 ∈ U the fast The main achievements of this period (1980s – system on the set bi(x) = bi(x0) converges to a unique 1990s) in model reduction were summarised in several stable fixed point x∗(b), where b is the vector of values books and surveys [13, 14, 15, 16]. bi(x). Then the slow system describes dynamics of con- The technological elaboration of these ideas and as- servation laws b: similation of those by the modelling practice took al- db ∗ most thirty years. Much efforts have been invested into = Db = ∗ [F (x (b))], (3) dt x x (b) slow computational improvements and testing with the sys- ∗ tems of various complexity. Some new ideas were pro- where Dbx=x∗(b) is differential of b(x) at the point x (b). posed and developed. For linear conservation laws Db = b and the slow equa- The QE, QSS, MIM, and CSP methods can be ap- tions have the simple form plied to any differential equation with explicit or im- db plicit (hidden) separation of time. They use the struc- = b[F (x∗(b))]. (4) dt slow ture of reaction network as a tool for creation of kinetic equations. In the classical methods, only the limiting The QE manifold is parametrised by the conservation step approach (the bottleneck method) works directly laws with functions x∗(b). It should be stressed that with the reaction graph. Recently, the model reduc- the slow equations in their natural form (3), (4) describe tion methods which use the structure of the reaction net- the dynamics of the conservation laws b and not the dy- work, were developed far enough and attract many dif- namics of the selected ‘slow coordinates’. The problem ferent techniques, from sensitivity analysis to algebraic of projection onto slow manifold is widely discussed geometry and tropical mathematics. [17, 18, 19]. According to the Tikhonov theorem, dy- The first step in the next section is a “step back- namics of the general system (1) from an initial state x0 wards”, a brief introduction of the classical methods. under the given assumptions can be split in two stages: Then we move to modern development. fast convergence to the QE manifold x∗(b) (the initial ∗ layer, convergence to a small vicinity of x (x0)), and then slow motion in a small vicinity of the QE mani- QE, QSS, MIM and CSP in ODE framework folds. The QE assumption is the separation of reactionsonto Formally, the standard models of chemical kinetics slow and fast: F includes all the terms from the slow are systems of Ordinary Differential Equations (ODE). slow reactions, F includes all the terms from the fast reac- The general framework looks as follows. Let U be fast tions and the slow manifold x∗(b) consists of equilibria a bounded domain in Rn. Assume that vector fields of the fast reactions parametrised by the conservation F x F x ff fast( ) and slow( ) are defined and di erentiable in laws. The ‘thermodynamic’ behaviour of fast reactions a vicinity of U¯ (in real applications these vector fields (convergence to equilibrium, which is unique for any are usually analytical, or even polynomial or rational). given values of the conservation laws) is essential to ap- U F x Let be positively invariant with respect to fast( ) and plication of the Tikhonov theorem. Slow reactions can F (x). Consider dynamical system with the explicit slow be extended by including external fluxes, they do not fast-slow time separation: change the asymptotic form (3), (4). dx 1 Combining of fast subsystems from the fast reactions = F (x) + F (x), (1) dt slow ε fast is so popular [22] that a special warning is needed: there 2 exists another widely used approximation without sepa- value of cg then we return to the previously described ration of reactions into fast and slow (see QSS below). situation: the motion can be separated into two steps: ∗ It should be stressed that the physical and chemical first, cs goes fast into a small vicinity of cs (cg) (ini- nature of the convergence to equilibrium of fast reac- tial layer) and then cg slowly changes in vicinity of this tions may vary. It may follow from thermodynamic manifold. conditions like principle of detailed balance or semi- In the QSS approximation, the smallness of amount detailed balance. It may have also completely alge- of intermediates is the crucial assumption. There may braic nature. For example, any linear (monomolecu- be various explanation of this smallness. For example, lar or pseudomonomolecular) system tends to a fixed in catalysis this is the smallness of the amount of the point, which is unique for any given values of con- catalyst. In combustion, the high activity of the inter- served quantities. The mass action law systems with mediates leads to their short life time. reactions ‘without interaction of different substances’ The list of variables for QSS exclusion can change α A → β A provide us with another example [25] in time. The smallness of concentration can serve as ri i P j r j j (here Ai are the components, r is the number of reaction, a criterion for the extension/reduction of this list [19]. and the coefficients are non-negative). The ‘speed coefficients’ were introduced recently [23] The QSS approximation does not separate the reac- for ranking variables according to how quickly their ap- tions into slow and fast. Usual application of the QSS proach their momentary steady-state. These coefficients is the slaving dynamics of the active intermediates in are used for a straightforward choice of variables for combustion (Semionov) or in catalysis (Hinshelwood). QSS elimination , while preserving dynamic character- The stable reagents participate in the same reactions as istics of the system. the intermediates do and, therefore, we cannon consider In addition to the different nature of small parameters, them as ‘slow’ reagents. The nature of small parame- the QSS differs from the QE in one more essential as- ter in this case is very different [20]. It is the small- pect: for the QE, the fast system is just chemical kinetics ness of the amount of intermediates. Let us illustrate of the set of fast reactions. It has all the thermodynamic this small parameter for heterogeneous catalytic reac- properties expected for the chemical kinetics equations, tions [21]. We use notations: V is the volume of reactor, like positivity of entropy production. It cannot have bi- S is the surface of catalyst, Ng is the vector of amounts furcations, oscillations, etc. The slow system in the QE of the gas components, Ns is the vector of amounts of approximation also has the thermodynamic properties. the surface components, cg = Ng/V, cs = Ns/S . Let us This is a particular case of the general theorem about the measure V and S in moles (i.e we use PV/RT for some preservation of the type of dynamics: the QE approx- average values of the pressure P and temperature T in- imation and the fast subsystem inherit thermodynamic stead of V and the number of moles of active centres on behaviour from the original system [24]. This property the surface instead of S ) for creation of dimensionless makes QE a promising initial approximation for various variables. The kinetic equations are: iterative model reduction procedures, [24]. In the QSS, the fast system can have many steady dNg dNs = SFg(cg, cs), = SFs(cg, cs), states, oscillation, etc. These critical effects in the QSS dt dt approach are considered as the symptoms of physical Where Fg and Fs are combined from the reaction rated critical effects, for example ignition: the original sys- per unit surface (mole of active centres). Notice that tem with thermodynamic behaviour has no bifurcations usually in the selected units S ≪ V (if the pressure is but just high sensitivity, at the same time the fast system not extremely low). Under given V, we can write equa- in the QSS approximation has bifurcations. S tions with a small parameter ε = V . If the fast system does not converge to equilibrium then the methods of averaging can work [28]. Instead dcg dcs = εFg(cg, cs), = Fs(cg, cs). of partial equilibria or quasi-steady-states the so-called dt dt ‘ergodes’ are used (subsets, on which the fast systems S are ergodic). After the change to the slow timescale, θ = V t we get the slow and fast subsystems: All the methods of invariant manifold in chemical kinetics were introduced for generalisation and improve- 1 c˙g = Fg(cg, cs);c ˙s = Fs(cg, cs). (5) ment of the QE and QSS. The earliest prototype was ε the Chapman–Enskog method in the kinetic foundation If the fast dynamicsof cs tends to the asymptotically sta- of fluid dynamics but this method used the Taylor se- ∗ ble fixed point (quasi-steady-state) cs (cg) for any fixed ries expansion in powers of a small parameter. The sec- 3 Fast equilibria duced. The method of kinetic polynomial gives the algebraic equation for the reaction rate [31]. This equation (the kinetic polynomial) is produced by the exclu- ఘଵ ଵ ఘଵ ଵ ߙ ܣ ߚ ܣ sion methods (the constructive algebraic geometry) and

ఘଶ ଶ ఘଶ ଶ is successfully used in analysis of catalytic reactions. ߙ ܣ ା ି ߚ ܣ ܤఘ ܤఘ The MIM aims to find the slow invariant manifold ,Fig. 2) [3, 4, 34, 19]. Invariance has simple definitions) ڭ ڭ ߙఘ௡ ܣ௡ Small amounts ߚఘ௡ ܣ௡ both analytic and geometric ones. Consider a system of ODE,x ˙ = F(x). Imagine that we found such a (nonlinear) coordinate transformation after that the vector x is Figure 1: The Michaelis–Menten–Stueckelberg limit: intermediates represented as a direct sum x = y ⊕ z and the equations B are in fast equilibrium with reagents Ai and are present in small amount (QSS). This limit results in mass action law for the brutto have the form reactions between A . i y˙ = Fy(y, z) (6) z˙ = Fz(y, z), ond approximation from this series is already singular. where Fy(0, z) = 0. This means that if y(0) = 0 then The iterative MIM was developed for the Boltzmann y(t) = 0 for solutions of the system. equations for correcting these singular properties of the The manifold y = 0 is invariant with respect to the Chapman–Enskog expansion [4, 26] and for solution of system. Of course, it is not necessary to assume a Hilbert’s 6th Problem [27]. The mathematically rigor- global transformation with this property. We can use ous version of the Chapman–Enskog method for finite- local transformations under condition that they define dimensional ODE was developed by Fenichel [32] and the same manifolds y = 0 in intersections (to provide was used as the prototype in many later works (see the gluing of the local pieces). The definition of slowness collection of papers [33]). is much more sophisticated. It can be done rigorously Combination of the QE and QSS assumptions (fast for systems with existing slow parameters ε using an- partial equilibria and small amounts of intermediates, alytical continuations from the very small values of ε for example) leads to important general results. It was or from a vicinity of an attractor (for example, a sta- used in 1913 by Michaelis and Menten. They proved ble fixed point) using separation of the relaxation modes that under this assumption the intermediates could be into fast and slow ones (for example, separation of the completely excluded from kinetic equations and the re- invariant subspaces of Jacobian by real parts of eigen- sulting kinetics will obey mass action law again. In values) [26]. 1952, Stueckelberg used this QE+QSS combination for Far from these continuations, there exist two popu- proving the semi-detailed balance condition for Boltz- lar heuristics: one used the eigenspaces of Jacobians mann’s equations (later, this condition was rediscov- in various states [16], another relies on a special it- ered and is popularised as the complex balance condi- eration procedures (a special version of the Newton– tion). The detailed review was presented and the gen- Krylov method) [13]. Both approaches converge to a eral Michaelis–Menten–Stueckelberg asymptotic theo- slow manifold for sufficiently small ε or in a vicinity of rem was formulated and proved in [29, 30] (see Fig. 1). an asymptotically stable equilibrium under some simple technical conditions (see [13, 35, 36]). If there exist no fast equilibria (no QE), then the QSS A special model reduction method, which utilised the approximation alone does not lead to the simple mass analytic continuation from a vicinity of a fixed point and action law after exclusion of the intermediates. Nev- Lyapunov auxiliary theorem was developed [37]. ertheless, if the mechanism of the intermediates transi- The CSP looks for more than just a slow manifold. It tions is linear then the explicit analytic expressions for tries to find the complete slow-fast decomposition, that the QSS reaction rates are obtained [21]. The simplest is to transform the initial system of ODEx ˙ = F(x) into of them, for the enzime (E) – substrate (S) reaction, a decoupled form S + E ⇆ SE → E + P, is known as the Michaelis– y˙ = Fy(y) Menten kinetics (produced by Briggs and Haldane in (7) 1925). z˙ = Fz(y), For non-linear reactions between intermediates, they where the non-diagonal terms (dependence ofy ˙ on z usually cannot be excluded from the equations analyt- andz ˙ on y) vanish. Of course, this complete decom- ically but an equation for the reaction rate can be pro- position could be rarely found explicitly (as well as the 4 Initial conditions

Initial Solution layer Projection to “slow variables”

Slow manifold

“Slow variables” Parameterization by “slow variables”

Figure 2: Slow Invariant Manifold slow invariant manifold) but the CSP tends to decrease The CSP was generalised and implemented as a part the non-diagonal terms iteratively. Again, the definition of a general strategy for analysis and reduction of un- of decomposition is simple and straightforward, but the certain chemical kinetic models [45]. attribution of slow/fast properties may be complicated Lumping analysis was reformulated as a particular and can be done rigorously for families of systems with case of MIM [46] and computational possibilities of this a small parameter or near a stable attractor. The classi- approach were tested. The reaction rate constants are cal heuristic for the slow-fast CSP decomposition used usually defined with uncertainty. A Bayesian automated the eigenvectors and eigenspaces of the Jacobians at all method was implemented for robust lumping for sys- states but very recently one of the inventors of the CSP tems with parameter variability [47]. The method works proposed a direct iteration method, which can lead to stepwise, reducing the system’s dimension by one at the same result avoiding expensive calculations of the each step. egienvalues and eigenvectors in transient states [40]. The theory and methods of invariant manifolds are During the last decade much work was done to im- used for modern development of the theory of criti- prove the computational efficiency and extend the area cal simplification. In 1944, Lev Landau noticed that of applicability of the MIM and CSP. In particular, near the loss of stability the amplitude of the emer- a grid-based computational method for slow invariant gent “principal motion” satisfies a very simple equa- manifolds was elaborated [41]. tion. It is an example of the “bifurcational parametric The method of Reaction-Diusion Manifolds simplification”. In chemical kinetics, The concept of (REDIM) [42, 43] was developed to extend the “critical” simplification was proposed in chemical ki- MIM for handling the diffusion processes in combus- netics by Yablonsky and Lazman (1996) for the oxidation problems [43]. It represents a modification of tion of carbon monoxide over a platinum catalyst using the Film Equation [13, 14] and of functional iteration a Langmuir-Hinshelwood mechanism. The main obser- technique [3] to approximate a slow invariant manifold vation was a simplification of the mechanism at igni- for a reaction-diffusion system. Recently, the REDIM tion and extinction points. This is a very general phe- method was improved to be able to generate and to nomenon known for various bifurcations. For the equa- approximate a slow manifold of arbitrary dimension by tions of chemical kinetics, the theory of critical simpli- employing its hierarchical structure [43]. The REDIM fication was developed recently [48] using the construc- was applied for model reduction in various combustion tive theory of invariant manifolds. systems . Various methods of construction of slow invariant 5 manifolds were tested and compared using a simple ex- tem change is slow and the domain where the change is ample [49]. Method of invariant grids [41] was em- relatively fast. Moreover, the fast area is expected to be ployed in this work for iteratively solving the invariance larger (the system ‘shrinks’ to the smaller-dimensional equation. Various initial approximations for the grid are manifold of slow motions). This heuristic idea gave rise considered such as QE, Spectral QE, ILDM and Sym- to development of a Singularly Perturbed Vector Fields metric Entropic ILDM (proposed earlier [19]). Slow in- (SPVF) theory [59] and of a Global Quasi-Linearization variant manifold was also computed using CSP method. (GQL) approach in its framework [60]. A comparison between method of invariant grids and In this way, the transformation to the explicitly de- CSP is also reported. Although CSP and the tested composed form (5) is constructed by a linear interpo- MIM are based on completely different construction of lation of the system vector field (1) from a randomized iterations, the comparison shows very similar results in sample of vectors x [59, 61, 62, 65]. The results of ap- terms of accuracy of slow invariant manifold descrip- plication of the GQL and the QSS were systematically tion. compared [63, 64]. Thermodynamics gives a convenient opportunity to The model reduction methods based on the ran- formulate the QE assumption. If the fast reactions domised sampling and principal component analysis are obey thermodynamics then the corresponding thermo- systematically used under the name ‘Proper Orthogonal dynamic potential is a Lyapunov function: it should Decomposition’ (POD) [66]. The POD is an a posteri- change monotonically along the fast motion [21, 13, 51, ori, data dependent projection method. The data points 53], and QE can be described as a conditionalextremum are either given by samplings from experiments or by (that is, maximum for entropies and free entropies or trajectories of the physical system extracted from sim- minimum for free energies) of the thermodynamics po- ulations of the full model (the so-called ‘snapshots’). tential. In this formalism the reactions are not needed, This point of view is often translated into the question: only the plane of fast motion is necessary. This sim- “Find a subspace approximating a given set of data in ple idea [19] gave rise to many ‘constrained equilibria’ an optimal least squares sense” [67]. Various linear and approximations [52, 50]. For example, a new version non-linear versions and generalisations of the principal of the QE was proposed, the spectral QE manifold, that component analysis are applied for dimensionality re- consists of partial equilibria in the planes of fast mo- ductions of snapshots [68]. In particular, a modifica- tions, which are parallel to fast eigenspaces of Jacobian tion of the POD with preservation of some original state at equilibrium [54]. variables is proposed [69]. Thermodynamics was used for comparing all re- We can expect intensive development of new model versible reactions by the same measure: entropy pro- reduction methods with applications of the classical and duction. It was proposed to exclude almost equilibrated new data mining techniques: independent component reversible reactions from the reaction mechanism [55]. analysis, various methods of clustering, artificial neu- The distance from the reaction equilibrium was mea- ral networks, and other machine learning methods. For sured by entropy production. The total entropy produc- example, the profile likelihood approach allows solving tion should not change significantly in this procedure. the model reduction problem together with the indetifi- (If we assume the detailed balance, then the entropy cation problem and analysis of parameter identifiability produced by a reversible reaction measures how far is and designate likely candidates for reduction. The fol- it from equilibrium.) This method was used for defining lowing references demonstrate some other recent efforts of the compartments in the composition space with the in this direction [71, 72, 73, 74, 75, 76, 77]. different reaction mechanism for hydrogen combustion [56]. Recently, this method was re-engineered and ap- Reaction networks, limiting steps, and dominant plied to auto-ignition of n-heptane with extraction of paths two skeletal mechanisms, and for analysis of spatially- varying premixed laminar flames [57]. For partially ir- Most of the methods described above can be applied reversible reactions, there exist special Lyapunov func- to more generalsystems of ODE than reaction networks. tions as well [58]. They exploit the slow-fast separation and some of them The fast-slow separation in the form (5) may be hid- utilise the thermodynamic Lyapunov functions (like QE den and should be revealed. Let us look on the sys- [19]). More rarely, the model reduction methods em- tem with fast-slow separation (Fig. 2). Two domains ploy both the thermodynamic functions and reaction can be distinguished there: the domain where the sys- mechanism, like, for example, the method for extraction 6 CAtt )( CAtt )( CAtt )( 1 2 q QS jick i A k ji Aj Aj j 1 Ai A C1 C2 Cq Ai k ji maxl{kli} C § · ¨ QS ¸ ¦cl 1 ¨ A C ¸ a) b) c) © l ¹

Figure 3: Hierarchical dominant path for linear system: (a) for each component retain the output reaction with maximal reaction rate: in each reaction ‘fork’ remains only dominant output and the reaction graph becomes a discrete dynamic system; (b) this discrete dynamic system converges to cycles (or fixed points); (c) cycles are glued in points, the output reactions are renormalised. A new reaction graph is constructed, where some vertices are glued. Then return to the step (a) for the new network. Iterate until all the trajectories of the discrete dynamical system converge to fixed points. This hierarchy of networks represents the hierarchical dominant path and provide us with asymptotic formulas for eigenvectors and eigenvalues of the initial network [78, 82, 79]. of skeletal mechanisms comparing entropy production of the kinetic matrix is close to the second constant in by various reversible reactions [57]. order. There exists a family of the model reduction meth- The simple and widely used idea of the limiting ods, which intensively use the structure of reaction net- steps was transformed into calculation of the hierarchi- works. The oldest, simplest, and, perhaps, the most used cal dominant paths (Fig. 3) for multiscale reaction net- method of model reduction in chemical kinetics is the works [79]. Consider linear reaction network Ai → A j method of limiting step – or bottleneck [78]. with reaction rate constants k ji. Assume that in each re- Consider systems obeying mass action law: the reaction ‘fork’ there exists a dominant reaction (Fig. 3 a): action rate r of the reaction i αρiAi → i βρiAi is if k ji = maxl{kli} then kqi/k ji < ε for some small pa- αi P P rρ = kρ c , where Ai are the components, index ρ rameter ε and any q , j. Delete all the non-dominant Qi i is a reaction number, αρi and βρi are non-negative stoi- reactions from the network. As a result, each compo- chiometric coefficients, kρ is the reaction rate constant, nent Ai will have only one outgoing reaction. Thus, the and ci is the concentration of Ai. The stoichiometric network is transformed into a discrete dynamical sys- vector of the reaction γρ has coordinates γρi = βρi − αρi. tems, where Ai are states and reactions are transitions. The kinetic equations (under constant volume) are Every motion in such a dynamical system is attracted by a cycle or a fixed point (Fig. 3 b). c˙ = X rργρ, This is not the end of the story. Find for every cy- ρ cle C the quasi stationary distribution: if we numerate where k is the reaction number. the components in the cycle in the order of reactions QS 1 1 A chain of linear irreversible reactions with non-zero A1,..., Ak then c = / j , where ki is the reaction i ki P k j rate constants A1 → A2 → ... → Am has m non-zero rate constant of the reaction Ai → ... from the cycle. eigenvalues of the kinetic equation matrix, which coin- Glue cycles into new vertices with the constants of out- QS cide with −ki, where ki (i = 1,..., m − 1) is the reac- going reactions multiplied by ci (Fig. 3 c). Iterate the tion rate constant of the reaction Ai → Ai+1. There- steps a-c of the analysis. The procedure will converge fore, the relaxation rate is determined by the smallest in the finite number of steps. We obtain a hierarchical constant. In a irreversible cycle A1 → A2 → ... → dominant path: first, reaction converges to cycles (or Am → A1 the smallest constant determines the station- fixed points), then, slower, to cycles of cycles, etc. It is ary reaction rate. Indeed. in a steady state all the re- proven [79] that the eigenvalues and eigenvectors of the action rates in such a cycle should be equal: kici = initial network are well approximated by this procedure k jc j = w. Therefore, in a steady state ci = w/k j, and for sufficiently small ε. This approach was successfully w = c / (1/k ). If0 < k < εk (i , 1) then applied to modelling the mechanisms of microRNA ac- Pi i P j j 1 i k c > w > k c /(1 + mε). If0 < ε ≪ 1/m then tion [80, 81] and other biochemical reactions [82] 1 Pi i 1 Pi i w ≈ k c . The relaxation time for an irreversible In order to find the dominant subsystems and paths 1 Pi i cycle with limiting step is inverse second reaction rate in nonlinear reaction networks, the methods of tropical constant [79]. Indeed, the simple calculations show that (max,+) algebras were employed [83, 84]. The notion for sufficiently small ε the minimal non-zero eigenvalue of tropical equilibration was elaborated that provides 7 approximate descriptions of the slow invariant mani- networks [88] and tested successfully on synthetic and folds. Compared to computationally expensive numeri- experimental data. cal algorithms such as the CSP, the tropical approach is Asymptotology [79] and tropical asymptotics [83, symbolic. It operates by the orders of magnitudeinstead 84] operate by the orders of smallness rather than by of precise values of the model parameters. The tropical computational accuracy and aim to find the limit ‘skele- methods provide identification of metastable regimes, ton’ of the reaction mechanism (under some additional defined as the low dimensional regions of the phase conditions like preservation of connectivity or persis- space close to which the dynamics is much slower com- tence). pared to the rest of the phase space. These metastable Simultaneously, a new technique for model reduction regimes depend on the network topology and on the or- was developed on the basis of the idea of controllable ders of magnitude of the kinetic parameters. accuracy and error propagation [89]. A geometric error The local asymptotics of the algebraic curves are de- propagation method creates a hierarchy of increasingly scribed by the tropical (or min-plus) systems of equa- simplified kinetic schemes containing only important tions on their dominant exponents following the New- chemical paths. ‘Importance’ is defined using preset ton polygon method. A necessary (tropical) condition accuracy targets and user-defined error-tolerance. The of being a solution is the coincidence of the dominant proposed technology operates with large and very large orders of growth of at least of two monomials of the reaction mechanisms (hundreds and thousands of reac- equation. These conditions provide a system of linear tions). The model reduction in this technology is intrin- inequalities in the dominant exponents which one can sically connected to model identification and validation handle by means of the knownalgorithms for linear pro- procedures: model reduction is an important component gramming. Thus, the problem of solving a system of of these operations. algebraic equations is reduced to a combinatorial one For example, the advanced Directed Relation of determining the dominant monomials. Since solv- Graph with Error Propagation and Sensitivity Anal- ing tropical systems is easier in general than solving ysis (DRGEPSA) method was developed [90]. Two algebraic systems, one can get a benefit in calculating skeletal mechanisms for n-decane were generated by asymptotics. this method from a detailed reaction mechanism for n- Extensive application of this method to biochemical alkanes. The detailed mechanism included 2115 species network models demonstrates that the number of dy- and 8157 reactions [90]. namical variables in the minimal models of large bio- Another reaction mechanism reduction method, chemical networks may be rather small and the number Simulation Error Minimization Connectivity Method of metastable regimes is sub-exponential in the num- (SEM-CM) produces several consistent mechanisms, ber of variables and equations [85]. The dynamics of which include the preselected important species [91]. It the network can be described as a sequence of jumps starts from the preselected important species and adds from one metastable regime to another. A geometrically strongly connected sets of species. These strongly con- computed connectivity graph restricts the set of possible nected sets are identified using the Jacobian (sensitivi- jumps. The graph theoretical symbolic preprocessing ties). This growth of the mechanism is controlled by the method significantly reduces computational complexity simulation accuracy and is terminated when the required of the analysis of reaction networks and computation of accuracy is achieved. The method was tested on the re- parameter regions for networks multistationarity [86]. action mechanism with 6874 reactions of 345 species. Model reduction is a necessary tool for model com- The reduced mechanism included 246 reactions of 47 position in chemistry and biochemistry of large reaction species, and numerical simulations became two orders networks [87]. The model should be reduced even be- faster [91]. fore its creation [19]. The full model can be imagined Various methods of reaction mechanism ‘surgery’ is but its identification and reliable verification is often im- now one of the hottest topics in reduction of large and possible. To obtain the model that works, we have to very large models in chemical engineering [57, 92, 93, reduce the full model before identification. More pre- 94, 95]. Surprisingly, the geometric details of this anal- cisely, model identification and model reduction should ysis seem to be very close to the geometric methods of be combined in one process. There are several attempts asymptotology and tropical asymptotic. of such an approach. For example, the combination For large and very large reaction systems, the reaction of model reduction and parameter estimation based on graph can be considered as an analogue of a geographic Rao-Blackwellised particle filters decomposition meth- map of a large area. For such large graphs, lumping is ods was proposed for non-linear dynamical biochemical close to zooming of maps. Zooming provides users by 8 Figure 4: Principle of semantic zooming exploited by NaviCell tool [97] for visualizing complex reaction networks in cancer biology. Semantic zooming can be considering as a simple application of lumping for gradual hiding the details of complex molecular mechanisms. a tool for work on various levels of model granularity References [96] and gives a possibility to study interaction between processes at different levels of the hierarchy. The princi- [1] Gorban AN, Yablonsky GS: Three Waves of Chemical Dy- ple of semantic zooming [97] was used for development namics. Math. Model. Nat. Phenom. 2015, 10(5):1–5. tools for navigations at different levels, similarly to ge- [2] Gorban AN, Yablonsky GS: Grasping Complexity. Comput. Math. Appl. 2013, 65(10):1421–1426. ological information systems [98] (Fig.4). [3] •Roussel MR, Fraser SJ: Geometry of the steady–state approximation: Perturbation and accelerated convergence methods. J. Chem. Phys. 1990, 93:1072–1081. Conclusion [4] •Gorban AN, Karlin IV: Method of invariant manifolds and regularization of acoustic spectra. Transport Theory and Stat. Phys. 1994, 23:559–632. Three components of model reduction methodology [5] •Maas U, Pope SB: Simplifying chemical kinetics: intrinsic are proved to be useful: low– dimensional manifolds in composition space. Combust. Flame 1992, 88, 239–264. • Universal approaches developed for general finite- [6] •Lam SH, Goussis DA: The CSP Method for Simplifying dimensional or even infinite dimensional systems Kinetics. International Journal of Chemical Kinetics 1994 26, 461–486. (slow invariant manifolds, geometric singular per- [7] •Rabitz H, Kramer M, Dacol D: Sensitivity analysis in chemi- turbation theory, etc.). For them, chemical dynam- cal kinetics. Ann. Rev. Phys. Chem. 1983, 34:419–461. ics is an important source of challenges and appli- [8] •Li G, Rabitz H: A general analysis of approximate lump- cations; ing in chemical kinetics. Chemical Engineering Science 1990, 45(4):977–1002. [9] Ranzi E, Cuoci A, Faravelli T, Frassoldati A, Migliavacca G, • Thermodynamic approaches, which utilize the ba- Pierucci S, Sommariva S: Chemical kinetics of biomass pyrol- sic physical and chemical structures; ysis Energy Fuels 2008, 22(6):4292–4300. [10] Nguyen TT, Teratani S, Tanaka R, Endo A, Hirao M: De- • Algebraic approaches based on analysis of reaction velopment of a Structure-Based Lumping Kinetic Model for Light Gas Oil Hydrodesulfurization. Energy Fuels 2017 mechanism. 31(5):5673–5681. [11] Abou-Jaoudé W, Thieffry D, Feret J: Formal derivation of Interaction and mutual enrichment of these techniques qualitative dynamical models from biochemical networks. will provide the future development of the efficient Biosystems 2016 149:70–112. model reduction for better computational performance [12] •Newman M, Barabasi AL, Watts DJ: The Structure and Dy- namics of Networks. Princeton University Press, Princeton, NJ; and deeper understanding of chemical dynamics. 2011. [13] Gorban AN, Karlin IV. Invariant Manifolds for Physical and Chemical Kinetics. Springer, Berlin – Heidelberg; 2005. Acknowledgement [14] Gorban AN, Karlin IV, Zinovyev AY: Constructive methods of invariant manifolds for kinetic problems. Phys. Reports 2004, Supported by the University of Leicester, UK, and 396:197–403. [15] •Turányi T, Tomlin AS: Analysis of Kinetic Reaction Mecha- the Ministry of education and science of Russia (Project nisms. Springer, Berlin – Heidelberg; 2014. No. 14.Y26.31.0022). [16] Goussis DA, Maas U: Model reduction for combustion chem- 9 istry. In Turbulent Combustion Modeling. Springer, Berlin – Approximate Slow Invariant Manifold method is carefully anal- Heidelberg 2011:193–220. ysed fo reactiondiffusion equations with slow and fast reaction [17] •Gear CW, Kaper TJ, Kevrekidis IG, Zagaris A: Projecting to kinetics. The second-order errors were precisely evaluated. The a slow manifold: Singularly perturbed systems and legacy results were also illustrated with a simple benchmark model. codes. SIAM Journal on Applied Dynamical Systems 2005, [37] Kazantzis N, Kravaris C, Syrou L: A new model reduction 4(3):711-732. method for nonlinear dynamical systems. Nonlinear Dyn. [18] Gorban AN, Karlin IV: Uniqueness of thermodynamic pro- 2010 59(1):183–194. jector and kinetic basis of molecular individualism. Physica [38] Goussis DA: Quasi steady state and partial equilibrium ap- A 2004, 336(3-4):391–432. proximations: their relation and their validity. Combust. [19] •Gorban AN, Karlin IV: Method of invariant manifold Theor. Model. 2012, 16(5):869–926. for chemical kinetics. Chemical Engineering Science 2003, [39] Goussis DA. Model reduction: When singular perturba- 58(21):4751-4768 tion analysis simplifies to partial equilibrium approxima- [20] Segel LA, Slemrod M: The quasi-steady-state assumption: A tion. Combust. Flame 2015, 162(4):1009–1018. case study in perturbation. SIAM Rev. 1989, 31:446–477. [40] • • Lam SH: An Efficient Implementation of Computational [21] Yablonskii GS, Bykov VI, Gorban AN, Elokhin VI: Kinetic Singular Perturbation. Combustion Science and Technology Models of Catalytic Reactions. Elsevier, Amsterdam 1991. 2018, 190(1):157–163. [22] Lee CH, Othmer HG: A multi-time-scale analysis of chemical An efficient formulation of computational singular perturbation reaction networks: I. Deterministic systems. J. Math. Biol. without calculation of Jacobian eigenvectors and eigenvalues is 2010 60(3):387–450. proposed and tested on simple example. [23] West S, Bridge LJ, White MR, Paszek P, Biktashev VN: [41] Gorban AN, Karlin IV, Zinovyev AY: Invariant grids for reac- A method of speed coefficients for biochemical model re- tion kinetics. Physica A 2004 333:106-54. duction applied to the NF-κB system. J. Math. Biol. 2015 [42] Bykov V, Maas U: The extension of the ILDM concept to 70(3):591–620. reaction-diffusion manifolds. Combust. Theor. Model. 2007 [24] Gorban AN, Karlin IV, Ilg P, Ottinger¨ HC: Corrections and 11(6):839–862. enhancements of quasi-equilibrium states. J. Non-Newtonian [43] •Neagos A, Bykov V, Maas U: Adaptive hierarchical con- Fluid Mech. 2001, 96:203–219. struction of reaction–diffusion manifolds for simplified [25] Gorban AN, Bykov VI, Yablonskii GS: Thermodynamic func- chemical kinetics. Proceedings of the Combustion Institute tion analogue for reactions proceeding without interaction of 2017 36(1):663–672. various substances. Chem. Eng. Sci. 1986, 41(11):2739–2745. The concept of slow manifold was successfully implemented for [26] ••Gorban AN, Karlin I: Hilbert’s 6th Problem: exact and reaction-diffusion equations. The method of Reaction–Diffusion approximate hydrodynamic manifolds for kinetic equa- Invariant Manifolds (REDIM) was developed tions. Bulletin of the American Mathematical Society 2014, [44] Strassacker C, Bykov V, Maas U: REDIM reduced modeling 51(2):187–246. of quenching at a cold wall including heterogeneous wall re- Method of slow invariant manifolds for kinetic equations is ex- actions. Int. J. Heat Fluid Flow 2018, 69:185–193. plained in detail with applications to Boltzmann’s equation. [45] •Galassia RM, Valorani M, Najm HN, Safta C, Khalil M, Ciot- [27] Slemrod M: From Boltzmann to Euler: Hilbert’s 6th prob- toli PP: Chemical model reduction under uncertainty. Com- lem revisited. Comput. Math. Appl. 2013 65(10):1497–1501. bust. Flame 2017, 179:242–252 [28] Slemrod M, Acharya A: Time-averaged coarse variables for CSP is employed to generate simplified kinetic mechanisms, multi-scale dynamics. Quarterly of applied mathematics starting from a detailed reference mechanism with uncertainty. 2012, 70(4):793–803. Uncertainty is modelled by random rate constants. [29] •Gorban AN, Shahzad M: The Michaelis-Menten- [46] Okeke BE, Roussel MR: An Invariant-Manifold Approach to Stueckelberg theorem. Entropy 2011, 13(5):966–1019. Lumping. Math. Model. Nat. Phenom. 2015, 10(3):149–67 [30] Gorban AN, Kolokoltsov VN: Generalized mass action law [47] Dokoumetzidis A, Aarons L: A method for robust model or- and thermodynamics of nonlinear Markov processes. Math. der reduction in pharmacokinetics. Journal of pharmacoki- Model. Nat. Phenom. 2015, 10(5):16–46. netics and pharmacodynamics 2009, 36(6):613. [31] •Marin G, Yablonsky GS: Kinetics of chemical reactions. De- [48] •Gol’dshtein V, Krapivnik N, Yablonsky GS: About bifurca- coding complexity. John Wiley & Sons, Weincheim, 201c1. tional parametric simplification. Math. Model. Nat. Phenom. A modern textbook in chemical kinetics, experimental and the- 2015, 10(3):168–185. oretical, fundamental and applied. In particular, in Chapter 9 the [49] •Chiavazzo E, Gorban AN, Karlin IV. Comparison of invari- theory of kinetic polynomials is presented. ant manifolds for model reduction in chemical kinetics. Com- [32] Fenichel N: Geometric singular perturbation theory for or- mun. Comput. Phys. 2007, 2(5):964-92. dinary differential equations. J. Diff. Eq. 1979, 31:59–93. [50] •Ren Z, Lu Z, Gao Y, Lu T, Hou L: A kinetics-based method [33] Gorban AN, Kazantzis N, Kevrekidis IG, Ottinger¨ HC, Theodor- for constraint selection in rate-controlled constrained equi- opoulos C (Eds.): Model Reduction and Coarse–Graining librium. Combust. Theor. Model. 2017, 21(2):159–182. Approaches for Multiscale Phenomena. Springer, Berlin- The rate-controlled constrained-equilibrium (RCCE) method is Heidelberg-New York 2006. compared to the classical QE approximation and it is demon- [34] Singh S, Powers JM, Paolucci S. On slow manifolds of chem- strated that for the systems with slow and fast reactions these ically reactive systems. J. Chem. Phys. 2002, 117(4):1482– methods are equivalent The detailed analysis allows authors to 1496. claim that the thermodynamics-based constrained equilibrium [35] Kaper HG, Kaper TJ, Zagaris A: Geometry of the computa- manifolds (CEMs) give a good approximation to the actual slow tional singular perturbation method. Math. Model. Nat. Phe- invariant manifolds in reactive systems. nom. 2015, 10(3):16–30. [51] Hangos KM: Engineering model reduction and entropy- [36] •Wu X, Kaper TJ: Analysis of the approximate slow invari- based Lyapunov functions in chemical reaction kinetics. En- ant manifold method for reactive flow equations. Journal of tropy 2010, 12(4):772–797. Mathematical Chemistry 2017, 55(9):1–30. [52] •Hiremath V, Pope SB: A study of the rate-controlled

10 constrained-equilibrium dimension reduction method and position. In Model Order Reduction: Theory, Research As- its different implementations. Combust. Theor. Model. 2013, pects and Applications Mathematics in Industry book series 17(2):260–293. (MATHINDUSTRY, volume 13). Springer, Berlin – Heidelberg [53] Hangos KM, Magyar A, Szederkényi G: Entropy-inspired 2008:95–109. Lyapunov functions and linear first integrals for posi- [68] Gorban AN, Zinovyev A: Principal manifolds and graphs in tive polynomial systems. Math. Model. Nat. Phenom. 2015, practice: from molecular biology to dynamical systems. Int. 10(3):105–123. J. Neural Syst. 2010, 20(3):219–232. [54] Kooshkbaghi M, Frouzakis CE, Boulouchos K, Karlin IV: Spec- [69] Chu Y, Serpas M, Hahn J: State-preserving nonlinear model tral quasi-equilibrium manifold for chemical kinetics. The reduction procedure. Chem. Eng. Sci. 2011 66(17):3907–3913. Journal of Physical Chemistry A. 2016 120(20):3406-3413. [70] Maiwald T, Hass H, Steiert B, Vanlier J, Engesser R, Raue A, [55] Bykov VI, Yablonskii GS, Akramov TA, Slin’ko MG: The rate Kipkeew F, Bock HH, Kaschek D, Kreutz C, Timmer J: Driv- of the free energy decrease in the course of the complex ing the Model to Its Limit: Profile Likelihood Based Model chemical reaction. Dokl. Akad. Nauk USSR 1977, 234(3):621– Reduction. PloS One 2016, 11(9):e0162366. 624. [71] Gorban AN, Zinovyev A Principal manifolds and graphs in [56] Dimitrov VI: Simple kinetics, Nauka, Novosibirsk, 1982. practice: from molecular biology to dynamical systems. In- [57] •Kooshkbaghi M, Frouzakis CE, Boulouchos K, Karlin IV: En- ternational journal of neural systems 2010, 20(03):219–232. tropy production analysis for mechanism reduction. Com- [72] Hiremath V, Ren Z, Pope SB: Combined dimension reduc- bust. Flame 2014, 161(6):1507-1515. tion and tabulation strategy using ISATRCCEGALI for the A method is developed for eliminating species from detailed re- efficient implementation of combustion chemistry. Combust. action mechanisms in order to generate skeletal schemes. The Flame 2011, 158(11):2113–2127. approach is based on the relative contribution of each elemen- [73] Amato F, González-Hernández JL, Havel J: Artificial neural tary reaction to the total entropy production. networks combined with experimental design: a “soft” ap- [58] •Gorban AN, Mirkes EM, Yablonsky GS: Thermodynam- proach for chemical kinetics. Talanta 2012, 93:72–78. ics in the limit of irreversible reactions. Physica A 2013, [74] Perini F: High-dimensional, unsupervised cell clustering for 392(6):1318–1335. computationally efficient engine simulations with detailed The systems with irreversible reactions are represented as the combustion chemistry. Fuel 2013, 106:344–356. limits of fully reversible systems when some of the equilibrium [75] Mirgolbabaei H, Echekki T: A novel principal component concentrations tend to zero. If the reversible systems obey the analysis-based acceleration scheme for LESODT: An a pri- principle of detailed balance then the limit system with some ori study. Combust. Flame 2013, 160(5):898–908. irreversible reactions must satisfy the extended principle of de- [76] •Zinovyev A: Overcoming Complexity of Biological Systems: tailed balance: from Data Analysis to Mathematical Modeling. Math. Model. (i) The reversible part satisfies the principle of detailed balance; Nat. Phenom. 2015, 10(3):186-205. (ii) The convex hull of the stoichiometric vectors of the irre- A detailed review about complexity reduction in biochemical versible reactions does not intersect the linear span of the stoi- systems. chiometric vectors of the reversible reactions. [77] Li S, Yang B, Qi F: Accelerate global sensitivity analysis using These conditions imply the existence of the global Lyapunov artificial neural network algorithm: Case studies for com- functionals. Thermodynamic theory of the irreversible limit of bustion kinetic model. Combust. Flame 2016, 168:53–64. reversible reactions is illustrated by the analysis of hydrogen [78] Gorban AN, Radulescu O: Dynamic and static limitation in combustion. reaction networks, revisited. Adv. Chem. Eng. 2008, 34103– [59] Bykov V, Goldfarb I, Gol’dshtein V: Singularly Perturbed Vec- 173. tor Fields, J. Phys.: Conf. Ser. 2006, 55:28–44. [79] • •Gorban AN, Radulescu O, Zinovyev AY: Asymptotology of [60] Bykov V, Gol’dshtein V: On a Decomposition of Motions and chemical reaction networks. Chem. Eng. Sci. 2010, 65:2310– Model Reduction. J. Phys.: Conf. Ser. 2008, 138:012003. 2324. [61] •Bykov V, Gol’dshtein V, Maas U: Simple global reduc- [80] Morozova N, Zinovyev A, Nonne N, Pritchard LL, Gorban AN, tion technique based on decomposition approach. Combust. Harel-Bellan A: Kinetic signatures of microRNA modes of Theor. Model. 2008, 12(2):89–405. action. RNA 2012, 18(9):1635–1655. [62] Bykov V, Griffiths JF, Piazesi R, Sazhin SS, Sazhina EM: The [81] •Zinovyev A, Morozova N, Gorban AN, Harel-Belan A: Math- application of the Global Quasi-Linearisation technique to ematical Modeling of microRNA-Mediated Mechanisms of the analysis of the cyclohexane/air mixture autoignition. Ap- Translation Repression. In MicroRNA Cancer Regulation: Ad- plied Mathematics and Computation 2013, 219(14):7338–7347. vanced Concepts, Bioinformatics and Systems Biology Tools, [63] •Yu C, Bykov V, Maas U. Global Quasi-Linearization (GQL) Advances in Experimental Medicine and Biology series, Vol. versus QSSA for a hydrogen-air auto-ignition problem. 774, Springer, 2013:189–224. Phys. Chem. Chem. Phys. 2018, DOI:10.1039/C7CP07213A. [82] Radulescu O, Gorban AN, Zinovyev A, Lilienbaum A: Robust Global quasi-linearization was explained and systematically simplifications of multiscale biochemical networks. BMC compared with QSS. Syst. Biol. 2008, 2(1):86. . [64] Nave O: Singularly Perturbed Vector Field Method (SPVF) [83] •Radulescu O, Vakulenko S, Grigoriev D: Model Reduction Applied to Combustion of Monodisperse Fuel Spray. Differ- of Biochemical Reactions Networks by Tropical Analysis ential Equations and Dynamical Systems 2017:1–18. Methods. Math. Model. Nat. Phenom. 2015, 10(3):124–138. [65] Bykov V, Gol’dshtein V: Fast and Slow Invariant Mani- Theory of tropical asymptotics (max,+) algebras and geometry folds for Chemical Kinetics.Comput. Math. Appl. 2013 65(10), of tropical equilibration of chemical reactions. 1502–1515. [84] Samal SS, Grigoriev D, Frhlich H, Weber A, Radulescu O: A [66] Willcox K, Peraire J: Balanced model reduction via the geometric method for model reduction of biochemical net- proper orthogonal decomposition. AIAA journal 2002, works with polynomial rate functions. Bull. Math. Biol. 2015, 40(11):2323-30. 77(12):210-2211. [67] Pinnau R: Model reduction via proper orthogonal decom- [85] Grigoriev D, Radulescu O, Samal S, Naldi A, Theret N, Weber

11 A: A Geometric analysis of pathways dynamics: application to versality of TGF-beta receptors. Biosystems 2016, 149:3– 14. [86] England M, Errami H, Grigoriev D, Radulescu O, Sturm T, We- ber A: Symbolic versus numerical computation and vizual- ization of parameter regions for multistationarity of biological networks. Proc. Intern. Symp. Comp. Algebr. Symb. Com- put., Beijing, 2017, Lect. Notes Comput. Sci. vol. 10490:93–108. [87] Kutumova E, Zinovyev A, Sharipov R, Kolpakov F. Model composition through model reduction: a combined model of CD95 and NF-B signaling pathways. BMC Syst. Biol. 2013, 7(1):13. [88] Sun X, Medvedovic M: Model reduction and parameter estimation of non-linear dynamical biochemical reaction networks. IET Syst. Biol. 2016, 10(1):10-16. [89] Pepiot-Desjardins P, Pitsch H: An efficient error-propagation- based reduction method for large chemical kinetic mechanisms. Combust. Flame 2008, 154(1):67-81. [90] •Niemeyer KE, Sung CJ, Raju MP: Skeletal mechanism generation for surrogate fuels using directed relation graph with error propagation and sensitivity analysis. Combust. Flame 2010, 157(9):1760–1770. [91] Nagy T, Turányi T: Reduction of very large reaction mechanisms using methods based on simulation error minimization. Combust. Flame 2009, 156(2):417–428. [92] Xin Y, Sheen DA, Wang H, Law CK: Skeletal reaction model generation, uncertainty quantification and minimization: Combustion of butane. Combust. Flame 2014, 161(12):3031– 3039. [93] Liu T, Jiaqiang E, Yang W, Hui A, Cai H: Development of a skeletal mechanism for biodiesel blend surrogates with varying fatty acid methyl esters proportion. Appl. Energy 2016, 162:278–288. [94] Gao X, Yang S, Sun W: A global pathway selection algorithm for the reduction of detailed chemical kinetic mechanisms. Combust. Flame 2016, 167:238–247. [95] Shakeri A, Mazaheri K, Owliya M: Using sensitivity analysis and gradual evaluation of ignition delay error to produce accurate low-cost skeletal mechanisms for oxidation of hy- drocarbon fuels under high-temperature conditions. Energy Fuels 2017, 31(10):11234–11252. [96] Sunnåker M, Cedersund G, Jirstrand M: A method for zooming of nonlinear models of biochemical systems. BMC Syst. Biol 2011 5(1):140. [97] •Kuperstein I, Cohen DP, Pook S, Viara E, Calzone L, Baril- lot E, Zinovyev A: NaviCell: a web-based environment for navigation, curation and maintenance of large molecular interaction maps. BMC Syst. Biol 2013 7(1):100. NaviCell Web Service (http://navicell.curie.fr) is a tool for network-based visualization of ‘omics’ data which im- plements several data visual representation methods and utilities for combining them together. It uses Google Maps and semantic zooming to browse large biological network maps with different types of the molecular data mapped on top of them. [98] •Dorel M, Viara E, Barillot E, Zinovyev A, Kuperstein I: Navi- Com: a web application to create interactive molecular network portraits using multi-level omics data. Database 2017, 2017(1):bax026. NaviCom, a Python package and web platform for visualization of multi-level omics data on top of biological network maps. NaviCom is bridging the gap between cBioPortal, the most used resource of large-scale cancer omics data and NaviCell, a data visualization web service that contains several molecular network map collections.