Development of façade fire incident database for a machine learning environment

Revision: R02 Date: 08/07/2019 Project Number: MA17068 Authors and reviewers: Michael Spearpoint, Ian Fu, OFR Consultants Kevin Frank, BRANZ Ltd, New Zealand Approved by: Simon Lay

BATH | EDINBURGH | GLASGOW | LEEDS | | MANCHESTER | OXFORD COMPANY NO. 9834368 | +44 (0)330 995 0648 | OFRCONSULTANTS.COM

Prepared for: Council on Tall Buildings and Urban Habitat

Prepared by: OFR Consultants Ltd Jactin House 24 Hood Street Manchester, M4 6WX

Project No: MA17068 Revision: R02 Date: 08/07/2019

The original version of this study was published 05/07/2018

© Olsson Fire & Risk UK Ltd All rights reserved. Olsson Fire & Risk UK Ltd, trading as OFR Consultants, has prepared this document for the sole use of the Client and for a specific purpose, each as expressly stated in the document. No other party should rely on this document without the prior written consent of OFR Consultants. OFR Consultants undertakes no duty, nor accepts any responsibility, to any third party who may rely upon or use this document. This document has been prepared based on the Client’s description of its requirements and OFR Consultants experience, having regard to assumptions that OFR Consultants can reasonably be expected to make in accordance with sound professional principles. OFR Consultants accepts no liability for information provided by the Client and other third parties used to prepare this document or as the basis of the analysis. Subject to the above conditions, this document may be transmitted, reproduced or disseminated only in its entirety.

BATH | EDINBURGH | GLASGOW | LEEDS | LONDON | MANCHESTER | OXFORD COMPANY NO. 9834368 | +44 (0)330 995 0648 | OFRCONSULTANTS.COM

Abstract A number of high-profile fires have recently occurred which involved the façades of tall buildings around the world. Such fires pose a life safety hazard to the building occupants, a hazard to people in neighbouring property, cause damage to the building, present a challenge for the fire service, affect the operation of the building after the event and result in major news items around the world that damages the reputation of tall buildings. To mitigate the risk of future fires it is important to understand the factors that have led to the events. The research has put together a database of high-rise building façade fires which includes the general characteristics of the buildings, information on the fire incident and the types of materials used to create the façade. Unfortunately, much of the data available in the public domain regarding the incidents is sparse which means the content of the database is not a comprehensive as would be ideal. Notwithstanding the limitations of the database, the work examines how it can be applied to the emerging field of machine learning which may lead to future findings that could not be readily discovered by manual methods. Using a machine learning procedure on the data has demonstrated the viability of using this technique. An assessment has been made of the impact of using different machine learning model configurations. A preliminary Severity classification has been used to evaluate the predictive capability of an optimised model when different building and façade types are present in a potential fire incident. A parallel manual analysis of the database has found several interesting results related to the age of buildings and if they had been subsequently refurbished, the building height and façade type in cases where casualties occurred in an incident, the frequency of incidents in different regions and some tentative findings on the reliability of sprinklers systems in façade fire incidents.

i

Table of Contents

1 Introduction ...... 1 1.1 Background...... 1 1.2 Tall buildings ...... 1 1.3 Façades ...... 2 1.3.1 General description ...... 2 1.3.2 Types of assembly ...... 3 1.4 Sprinklers ...... 4 1.5 Meteorological conditions ...... 4 2 Database ...... 5 2.1 Development ...... 5 2.2 Incidents ...... 6 2.3 Basic building information ...... 6 2.4 Building floors and height ...... 8 2.5 Year built, refurbished ...... 10 2.6 Façade construction ...... 13 2.7 Cause, origin of fire and age of building when incident occurred ...... 15 2.8 Casualties...... 17 2.9 Sprinklers ...... 20 2.10 Wind ...... 21 2.11 Falling debris ...... 22 2.12 Analysis ...... 23 3 Machine Learning ...... 25 3.1 Software specification ...... 25 3.2 TensorFlow application ...... 26 3.3 Classification model trained on the surveyed dataset ...... 27 3.3.1 Overview ...... 27 3.3.2 Input data preparation ...... 28 3.3.3 Features and label ...... 29 3.3.4 Machine learning model ...... 30 3.3.5 Selection of optimal model parameters ...... 33 3.3.6 Results of optimised model and observations ...... 34 4 Conclusions and future work ...... 36 ii

5 Acknowledgements ...... 37 6 References ...... 37 Appendix A : Terminology ...... 39 Appendix B : Incident data...... 40 Appendix C : Model evaluated accuracy ...... 101 Appendix D: Detailed trained model information ...... 106 Appendix E: Predicted Severity classification scores of the best-performing optimised model ...... 107

iii

1 INTRODUCTION

1.1 Background Fires which involve the façade of a building of any size or type have the potential to result in casualties and property damage. When such a fire occurs on the façade of a tall building, this can lead to catastrophic outcomes as recently illustrated by the incident [1]. The potential for the development of fire on building façades has been recognised for several decades and so has been a topic of ongoing research. For example, there is extensive work in the literature looking at the dynamics of façade fires including the work by Delichatsios [2] (and co-workers). In the UK, work has been undertaken by BRE Global [3, 4] partially in response to past fire incidents involving building façades. In New Zealand BRANZ has carried out research into façade fires from the late 1980s onwards [5]. Several reviews of façade fires are available in the literature, particularly White and Delichatsios [7] and the earlier study by Wade and Clampett [6]. However, there appears to have been no systematic approach to determine what factors are leading to these façade fires and why the consequences between fires differ. Reasons might simply be due to the type of panel or the type of insulation material being used in isolation but it could be due to a combination of design, material, building, installation, incident and regulatory factors that may not be so obvious from a cursory inspection. The realisation that façade fires have the potential to cause unwanted consequences has led to the development of tests such as NFPA 285 [8] in the US, BS 8414 [9] in the UK, ISO-13785 [10] and more recently FM 4880 [11]. These standardised assessment methods vary in the types of façade assembly configurations, the fire exposure conditions, the data that is recorded and how the results are assessed to provide some form of rating and/or pass-fail criterion. Under certain circumstances results from standard fire tests can be extended beyond the specific test conditions, where appropriate, using expert judgement. An example in the UK is the use of what is referred to as a BRE 135 [12] desktop extended application assessment. Regulatory mechanisms are often put in place by jurisdictions to control the type of façade systems in relation to the performance against standard tests, the type of building use, other fire protection measures, etc. These controls may be as a result of previous incidents, and for example, in response to several high-profile incidents the UAE revised its building code in 2013 with regard to the fire performance of cladding on all new buildings over 15 m tall. Regardless of the past research, available assessment pathways and regulatory approaches, fires that have involved façades on tall buildings appear to occur more frequently with much more dramatic outcomes than in the past. One reason for this might be that tall buildings have become more common in cities, for example statistics compiled by the CTBUH show that the number of buildings of 200 m height or greater completed in 2016 continued to increase compared with previous years [13]. However, it seems that factors such as the materials, design and installation methods used to construct façades have changed so as to increase the fire risk. This exploratory research is the result of seed funding through the Council of Tall Buildings and Urban Housing (CTBUH) with the objective identify existing reviews of façade fires on tall buildings, to create a database of relevant façade fire incidents and to examine where the emerging field of machine- learning might be applied to the analysis of the data. A view of the database of incidents created by this work is given in Appendix B. Each incident is given an identification (ID) number and where incidents are specifically discussed in this report then a cross-reference to its ID is provided. 1.2 Tall buildings The focus of this work is on façade fire incidents involving ‘tall’ buildings with no specific occupancy type targeted. Thus, single use buildings such as hotels, residential or office and multi-use buildings 1

are included where such buildings have experienced a relevant fire incident. The CTBUH notes that “there is no absolute definition of what constitutes a ‘tall building’; the definition is subjective...” but “…a building of 14 or more stories – or more than 50 meters in height – could typically be used as a threshold for a tall building” and so this definition is generally adopted here. However, façade fires of interest have occurred in buildings that do not meet the minimum definition and so these have also been included in this work. Façade fires in ‘low-rise’ buildings, such as single storey industrial units and single-family residences, are not included in this study although many such incidents were identified during the work. Thus, the term ‘high-rise building’ is generally applied in this report to collectively include tall buildings and those others that do not meet the CTBUH threshold. 1.3 Façades 1.3.1 General description A façade is used to form the exterior of a building. It sets the architectural tone of the building but also impacts on key design considerations such as: the energy efficiency, weathertightness, etc. Typically, façade systems consist of an assembly of components that are suspended or attached to the building structure, plus other associated external building elements such as: windows, balconies, etc. The building structure usually consists of precast concrete slabs or a steel frame. The terminology surrounding façades is not wholly consistent within the literature and so a particular vocabulary is adopted in this study as described in Appendix A. It is not the purpose of this study to present a detailed description of façade system design and a useful summary is available in White and Delichatsios [7]. In broad terms, and in the context of this study, a façade assembly can be either a non-structural curtain wall system or consist of an external cladding that is fixed to the solid structure of the building. Curtain walls can be made of glass or other lightweight materials to provide weathertightness, etc. Cladding may be made from panels or contiguous layers of material either mechanically fastened or attached to the building structure with an adhesive. A façade system can either include ventilated cavities or be considered as unventilated. Double-skin curtain walls allow for natural or mechanically-driven air flow within an intermediate cavity. If glass skins are used they can be double or single glazed and may have solar shading devices placed within the cavity. Façade systems that are not ventilated may be formed using a render or through the attachment of items such as wall panels to the building structure. Although the means of attaching panels may result in a relatively small gap between the structure and the panel, this is not treated as being a ventilated cavity. Façade systems with cavities may include barriers to prevent fire and smoke spread within the cavity space. These barriers typically consist of non-combustible components or intumescent materials. The panels or render components used as a cladding often act as an insulation barrier to increase the thermal comfort of the building. From a strict heat transfer point of view, curtain wall components also act a thermal insulator. However, due to their relative thinness, in terms of building comfort they serve little purpose. Therefore, curtain walls may include an insulation layer attached to the building structure and this layer may form part of the inner surface of a cavity. Insulation materials typically include mineral fibres, polyurethane (PU) foam, foamed phenolic, polyisocyanurate (PIR), expanded polystyrene (EPS) or extruded polystyrene (XPS). In addition to insulation materials, a façade assembly may include a water resistive layer. This generally consists of a thin membrane that prevents moisture travelling through a façade assembly and/or into the internal building environment itself. Materials include sheets of polyethylene or polyester and in terms of the overall fire performance of a façade are likely to have a minor impact when compared to layers of insulation and/or curtain wall panels. Structural membranes can be used as a lightweight building façade option [14] but these are beyond the scope of this study.

2

In some buildings the façades consist of a homogeneous assembly on one or more elevations. However, many buildings, and in particular those with apartments, will likely have balconies which disrupt a façade assembly. These balconies will provide a pathway for a façade fire to enter a building. They have the potential to affect the flame spread dynamics through a combination of horizontal and vertical surface orientations. It is also possible the balconies are constructed with materials that differ from the façade assembly. For example, Holland et al. [3] partly attribute the extent of fire spread in a 1960’s residential building to combustible insulating panels clad onto the underside of all the building’s balconies. Balconies also likely increase the probability of ignition due to human activities such as cooking, smoking, etc. Façade systems will also include specific design details such as fire stopping around penetrations through the façade assembly, e.g. windows, etc. The relationship between a façade assembly and windows appears to be a topic for which there seems to be less consideration when compared to the fire performance of the façade assemblies themselves. Glazing assemblies may be single pane or multiple pane, glass may be fire resistant and there may be openable windows which could be manually or automatically controlled. In the event of a façade fire, having openable windows could allow for spread into the building. In addition to the glazing, an assembly will include a window frame made from materials such as uPVC, timber or aluminium, plus the associated window hardware such as handles, locks, hinges, etc. The performance of glazing subject to flame radiation has been investigated by several researchers and Babrauskas [15] gives a useful summary with recent work reported by Holland et al. [3] specifically investigating incidents involving double glazing and external walls. The breaking of glazing provides another mechanism for fire spread between a façade fire and the inside of the building and recent work by Wang et al. [16] has investigated the fracture behaviour of glazed curtain walls related to this matter. 1.3.2 Types of assembly One common type of façade assembly uses metal composite material (MCM) panels as part of its construction. The MCM panel component is made up of two layers of metal skin with a core material in-between. The metal skins may be surface powder-coated or anodized aluminium, stainless steel or titanium. Core materials include polyethylene (PE), polypropylene (PP) or a fire-retardant formulation consisting of a thermoplastic with typically around 70 % mineral fill added. A common example of a type of panel used in MCM assemblies is aluminium composite material (ACM). These panels are then suspended from the building structure by fixing them to horizontal and/or vertical rails. Another typical assembly used in façade systems are Exterior Insulation Finish Systems (EIFS), also referred to as External Thermal Insulation Composite Systems (ETICS). These were developed in the 1950s [17] and use a layer of insulating material such as: expanded polystyrene, polyurethane, or polyisocyanurate on a non-combustible substrate and one or more thin outer finish layers that may include a reinforcing mesh layer and coatings. In this work façade assemblies are considered in terms of their material characteristics since a specific component can be used in different ways depending on the façade build-up. The term component is used here to refer to a material or a combination of materials that form a distinct item, for example an ACM panel. As such, a component may form part of a ventilated cavity so that it adjoins an air gap or could be in direct contact with an insulation material. Thus, it is the layers of material types (and any associated air gaps) and their thicknesses that are likely to influence the fire performance of a façade system rather than the specific components in isolation. Materials have several relevant thermo-physical / kinetic properties including: density, thermal conductivity, specific heat, heat of combustion and activation energy. Some of these are functions of temperature and obtaining such properties can be challenging. Measurement may be difficult, tests may not have been carried out to measure a specific property or tests have been conducted but the results have not been released. Even where tests have been completed and results made available the extent of the information available may be restricted due to commercial reasons. 3

1.4 Suppression systems There has been considerable recent discussion in the media about the role that suppression systems might play in the mitigation of façade fire incidents. Suppression systems can include traditional sprinkler systems, water mist systems, gas flood systems, etc. Sprinklers systems can be specifically designed for commercial applications or for residential buildings and are typically installed within the building envelope but can be applied as an external building system. In the context of façade fire incidents, particular focus has been on sprinkler systems installed inside the building envelope. As already noted by White and Delichatsios [7] several jurisdictions require or recommend internal sprinkler systems. It is quite possible that an internal sprinkler system will have little or no effect on fires that only involve the façade. A sprinkler system could aid in mitigating internal fire spread should flames enter the building through broken windows or gaps, for example. However, it is possible that a sprinkler system could become overwhelmed should the façade fire enter the building at multiple locations. Such an event could challenge a sprinkler system by exceeding the flow and pressure design criteria. Where a sprinkler system may be effective is in incidents in which the fire starts internally with the potential to spread to the façade through openings already present prior to the incident or created during the incident. The successful intervention of a sprinkler system would likely mean the façade never becomes involved. Any proposal to install a fire safety system should assess its reliability as part of an overall cost-benefit analysis. This is not only relevant to sprinkler systems but also for any other active or passive system. Statistics collated by Frank et al. [18] show values for the effectiveness of sprinkler systems on internal building fires in the range of 70 % to 99.5 %. 1.5 Meteorological conditions The climatic conditions that a building experiences will impact on the choice of a façade system, its long-term performance and how it may react during a fire. Over longer timescales the climate can affect the performance of materials. UV from sunlight can age materials, moisture can cause swelling, contribute to the rotting of organic materials and be instrumental in rusting iron-based materials, etc. Of the various potential influences of weather on façade fires the effects of wind are an obvious factor to consider. Compared to low-rise structures, taller buildings are subjected to greater wind speeds generally since they are higher in the boundary layer and are not often sheltered by other buildings/geography. Wind may enhance the spread of flame across surfaces and may increase the likelihood of burning pieces of material being dislodged. Wind may blow flames towards or away from the building that is burning or neighbouring buildings. Similarly, wind may influence the movement of smoke around the burning building and near neighbouring structures. The seasonal outside ambient temperature conditions will affect the selection of façade system and its relationship with the building heating and ventilation systems. For example, whether a façade system includes insulation and to what degree, also the effect of solar gain on ventilation requirements. The ambient temperature may have an effect during a fire since flames spread more rapidly over warmer materials as less energy is required to raise the material temperature to its ignition point. Other metrological factors could include the barometric pressure. Research has shown that this can also affect flame spread rate [19]. A beneficial effect of weather could be where precipitation on a façade fire might contribute to fire suppression, although it would be difficult to justify reliance on such a circumstance from a design or operational perspective. None of the incidents identified in this work showed any fire mitigation benefits of precipitation albeit few incidents occurred where it was raining at the time.

4

2 DATABASE

2.1 Development Some previous work has been carried out to collate information from real fire incidents in high-rise buildings, such as that by Wade and Clampett [6] and White and Delichatsios [7]. However, neither of these studies put together a data collection process that included a defined set of fields that could be systematically populated. One key objective of this work is to propose a set of data fields that are applicable to high-rise building façade fire incidents and then to populate those fields with as many incidents that can be identified. During the initial stages of the database development it was recognised that having data that described the general parameters associated with the building needed to be captured. Thus, information regarding the building height/number of storeys above ground; construction material; location; regulatory environment; years of design, completion and/or renovation; and the presence of a sprinkler system were all considered to be relevant and likely to be obtainable. For this work it is clear that a database needs a description of the façade system to some level of detail. As already discussed above, the details associated with a façade system can be relatively complex even where a single system has been used for a building. However, it is possible that more than one system could be relevant to a single building. Unfortunately, obtaining a high level of detail of the façade systems used on the buildings in the incidents assessed has been a major challenge in this work. Reports often give very little information on the type of façade system, the components used and/or the development of the fire in relation to the façade. A similar set of issues was noted in the work of White and Delichatsios [7]. In addition to the information regarding the building and the façade system, it is important to identify specific parameters related to the incident. It is recognised that there is potentially a large volume of information that can be associated with a specific incident. As an example, the investigation of the Lacrosse Tower incident (ID 34) resulted in a 124-page detailed report by the Fire Brigade [20]. Key incident parameters related to a façade fire are: the cause of the fire, where fire started in relation to the façade (for example whether the fire originally started inside the building before spreading to the façade or whether the incident was the result of an external fire such as the burning of rubbish, etc.), on what floor the façade initially became involved, the mechanism of fire spread such as over the outer surface and/or within a cavity, the effect of any wind, and the intervention by manual (i.e. fire service) or automatic systems. This is not an exhaustive list and, for example, there are a number of prevailing climatic conditions such the specific wind speed and direction, the presence of rain, ambient temperature and humidity, etc. that could be added. However, the subsequent review of the incident information showed that such climatic data are generally unavailable although it may be possible to use national meteorological records to obtain this given the time and location of the incidents are generally known. In putting together any database there needs to be consideration of what it should or could include. Specifying a large range of parameters does not in any way guarantee that the fields are of sufficient relevance or that data is available to populate the fields within a database. There also needs to be consideration on whether data is best provided numerically, through specific categories or as free- form text, each with their own distinct advantages and disadvantages. The following sections describe the data fields in more detail. Several of the fields provide the data numerically. However, given the nature of the information found in the literature it is found that other fields were better expressed through the use of descriptive text. The mixture of numerical and textual fields is referred to in this study as the ‘primary’ database. However, to apply the data to the machine learning environment the descriptive text fields have been translated into categories and these are also discussed in the sections below. For each of the data fields, one category option available for the machine learning environment

5

is ‘Unknown’. The machine learning environment subset of the primary database is referred to as the machine learning database in this study. 2.2 Incidents Currently the primary database holds 59 fire incidents (where 1 incident is a collection of 4 fires on a particular housing estate) that meet the research objective. Four key references [6, 7, 17, 21] were initially used as a first pass to identify the frequently cited façade incidents. Subsequently other publications from the literature including various web-based resources such as from the news media were employed to gather further details or to add incidents not reported in the key references. However, it is important to note that the content of the key references often relied on news media reports as their primary sources. Some of the incidents have occurred subsequent to the publication of the key references or appear to be less well known. The oldest incident in the database is 393 Kennedy St, Canada (ID 20) and the most recent event is the Zen Tower, incident (ID 59) which occurred on the 14th May 2018. Finally, the CTBUH global tall building database (www. skyscrapercenter.com) was used to provide further information about the buildings, where available. Some additional general building information was also sourced from the Emporis (www.emporis.com) website. The various source materials consulted on a given incident do not always agree on the details which inevitably results in a level of uncertainty with the data. Where the details vary then decisions have been made by the authors of this work on what information is suitable to use to populate the database. The following sections discuss specific cases in which incident information conflicted and what decisions are made. Further investigation may reveal the need to update the database as appropriate and it is appreciated that other investigators may have made different decisions as to what information could be relied upon. Two buildings within the dataset have experienced two fires. Of the two, the fires in The Torch (ID 2 and ID 3) are probably the most well-known whereas the fact that two fires occurred in the Sulafa Tower (ID 27 and ID 28) is probably less appreciated. Even though the two Sulafa Tower incidents occurred four years apart, this partly might be due to the fact that the fires reportedly occurred on the 35th and 36th floors so to the casual reader it might be thought that there was some confusion about on which floor the fire/s occurred or some of the reports may have not have correctly cited the floor level. 2.3 Basic building information Collection of basic information regarding the building has been generally straightforward. The location has, where possible, identified the name of the building, the city or region and the country. However, occasionally names vary due to translation to English from the native language of the home country of the building. In addition, at least one building has changed its name since the incident occurred, where the Country Comfort Motel, Albury, NSW (ID 32) is now known as The Marque. In addition to the name and location of the building involved in the incident, the longitude, latitude and elevation above sea level coordinates are recorded in the database where possible. The incidents come from 21 countries (where those incidents specifically identified as being in England and Scotland have been combined with incidents only as being known to have occurred in the UK). Figure 1 shows the number of incidents from each country with the differentiation between UK / England / Scotland shown with the hatching. Each country bar is colour-coded by region, where the regions are those defined in the CTBUH database. It is important not to read too much into this data as there are a number of factors that come into play. Not only do the data include some incidents in which the fire was not severe but the extent of reporting and access to resources differs by country. Individual countries have their own specific population and building demographics and have many factors that may affect the number and types of fires that occur. Furthermore, the identification of incidents is likely influenced by the lead author being from a Western European country with English 6

as their primary language. For the machine learning environment, the country categories are as those shown in Figure 1.

12

10

8

6

4

2

0

USA

UAE

Israel

China

Qatar

Japan

Russia

France

Turkey

Canada

Hungary

Romania

Australia Moldova

Germany

Indonesia

Azerbaijan

Switzerland

South KoreaSouth UK / England / Scotland England UK / /

Figure 1 – Number of façade fire incidents by country. Regional colour codes: Asia – red; – green; Middle East – blue; Oceania – purple; North America – yellow.

The primary database identifies the use of the building, where some buildings had a single use whilst others had a mixture of uses. Typical uses include: residential apartments, hotels, office and parking with a few cases of retail, casino and medical occupancies. For the machine learning environment, the type of building is categorised into six occupancy use types as shown in Table 1 with the number and associated percentages indicated. Categories that allow for mixed use could have been defined but would have meant there would have been fewer records in each category increasing the data sparsity. Thus, buildings with a mixture of occupancy types are assigned a category that reflects the predominant use and/or where the incident primarily occurred. For example, in the Tamweel Tower, Dubai incident (ID 4) the building was a mixture of office and residential uses but the fire occurred in the residential part so this is assigned to the residential building type. As indicated by Table 1, the majority of incidents (76 %) occurred in residential (or the residential portion) buildings. Collectively 90 % of the incidents occurred in buildings that present a sleeping risk occupancy (residential, hotel, medical).

7

Table 1 – Building use type: Machine learning environment categories.

Category Description Number (%)

Residential Apartments, flats 45 (76 %)

Hotel Hotel, motel 7 (12 %)

Office - 3 (5 %)

Commercial Retail, leisure 3 (5 %)

Medical Hospital, medical centre 1 (2 %)

The primary database includes the main material of building construction. Around one third (34 %) of the buildings were constructed of (reinforced) concrete, a few (5 %) are identified as being steel and the remainder are of unknown construction. Thus, for the machine learning environment only three categories are defined: 'concrete', 'steel' and 'unknown'. More detailed information regarding the characteristics of the buildings have not been specifically identified in the database. For example, events such as the 2014 Lacrosse Tower (ID 34) and the 2010 La Tour Mermoz (ID 6) fire incidents had buildings that included balconies but buildings that had balconies are not distinguished from those without balconies in the current version of the primary database. It was hoped to include in the primary database some details on the local building regulations and design approach employed during the development of each building. Unfortunately, very little specific information was identified in the research, meaning adding such fields to the primary database was superfluous, at least until further details could be sourced. 2.4 Building floors and height The number of floors in the building (and the floor level where a fire started have been recorded). However, these data are not always consistent across the various references. In terms of the number of floors this may be due to the counting of any floors below ground; whether ground floor is included in the tally; whether split floors are counted as one or two levels; and/or whether the number of floors in the description of a building has been rounded to a specified number. It may also be that there has been some confusion regarding the name of the building when information has been reported in the media. For example, reports on the Central Television HQ and Mandarin Oriental Hotel, (12) incident quote the number of storeys as 31, 34 or 44. The CTBUH database lists the building as 31 storeys but at least two other reports give 44 storeys. However, it would appear that those reports are referring to the CCTV Headquarters which is listed as having 44 storeys according to Wikipedia (or 54 storeys according to the CTBUH database). A value of 31 storeys is applied to the database consistent with the CTBUH database. The CTBUH notes that the above ground number of floors includes the ground floor and main floors counting any significant mezzanine floors and major mechanical plant floors. The CTBUH also recognises that floor counts in other published accounts may differ because it is common for certain floor levels not to be included in some regions (e.g. levels 4, 14, 24, etc. in or level 13 in some other countries). The number of storeys have been reported for 55 out of 59 building incidents and storeys have been reasonably estimated for the four unknown buildings by counting likely floor levels using available photographs or from Google Street View. Figure 2 shows the cumulative number

8

of buildings in the database with N stories or less. Of the 59 incidents, 19 (~32 %) have 13 storeys or less and so do not generally meet the CTBUH definition of a tall building.

90

80

70

60

50

40

30 Number of stories above ground, N ground, above stories of Number 20

10

0 1 2 5 8 9 12 15 16 19 20 21 23 24 26 27 30 31 32 34 35 36 37 39 41 43 44 46 49 50 52 53 54 56 58 Number of incidents in buildings with N storeys above ground or less

Figure 2 – Cumulative number of façade fire incidents in buildings with N stories or less.

The primary database also includes the height of the building, where possible following the definition employed by the CTBUH as that measured “from the level of the lowest, significant, open-air, pedestrian entrance to the architectural top of the building, including spires, but not including antennae, signage, flag poles or other functional-technical equipment”. In some cases it is not clear what height is being referred to when it is stated in a report. The tallest building referred to in the database is the 352 m tall, 86 storey The Torch in Dubai whereas the shortest is a 5 storey residential building in . Figure 3 shows the height of each building plotted against the number of stories, where both data are known. Unsurprisingly there is a fairly direct correlation between the two data sets. The data point shown at 63 stories and 302 m is The Address (ID 1) where if the occupied building height of 228.3 m were applied this point would more closely fall in line with the other data. Since there is a direct correlation between building height and the number of storeys above ground the machine learning environment only applies the number of storeys as a numerical value.

9

400

350 Asia Europe 300 Middle East 250 North America 200 Oceania

150 Buildingheight, m 100

50

0 0 10 20 30 40 50 60 70 80 90 Number of storeys above ground

Figure 3 – Building height versus number of stories above ground.

2.5 Year completed / refurbished The data from the reports consulted is not always consistent as the definition of when a building is built may vary. The CTBUH considers a building to be complete when it has been: 1) topped out structurally and architecturally 2) fully-clad, and 3) open for business, or at least partially occupiable. However, it is likely these three criteria are not used universally and although this work uses the term ‘completion’, other sources use the term ‘constructed’. The oldest building dates from 1961 (Lomond House, ID 41) although there are other UK buildings from the 1960s included in the database in which the exact date of construction is not precisely known. In some of the incidents the building was still under construction. If this was the case this is separately identified in the database and the year of construction is listed as ‘n/a’ (not applicable). Several buildings had been modified prior to the incident. Typically, these changes have been made to older buildings and in some cases it is because of changes to the façade that the incident likely developed in the manner it did. The CTBUH tall building database differentiates between a building which has been renovated and one that has been retrofitted. Renovation is where a major alteration has occurred some time after the original construction had been completed which resulted in a significant change of height, appearance, and/or use. A building that has been retrofitted is one where extensive work has been carried out but there was no change in original appearance or height. The CTBUH façade retrofit database (facaderetrofit.org/) also provides definitions for renovation, retrofitting and a number of related terms. In this context renovation and retrofit differ from the tall building database. In this work the term ‘refurbished’ has been used to encompass both renovation and retrofit although, again, this definition differs from that in the façade retrofit database. For the machine learning environment, the year of construction is provided as a numerical value with a mid-decade estimate made for those buildings in which information has not been sourced. Similarly, the year of any retrofitting is given a numerical value, where appropriate. Figure 4 shows the year of building construction as a function of building height.

10

90 Asia 80 Europe 70 Middle East 60 North America 50 Oceania 40 30 20

10 Number Number of storeysabove ground 0 1960 1970 1980 1990 2000 2010 2020 Year of completion

Figure 4 – Year of building completion as a function of number of storeys and region.

What is clear from this is that very few incidents have occurred in buildings completed between the mid-1970s and the mid-2000s. There are several reasons why this might be the case. The first of these reasons likely includes the fact that, other than in North America, the rate of tall building completions through the 1980’s was relatively low worldwide. This can be illustrated by examining the CTBUH Tall Building database for buildings over 150 m tall completed each year in different regions over the past six decades. Figure 5 shows that in Europe tall building construction had reduced by the 1980’s compared with the earlier decade whereas few tall buildings had been completed in Asia and the Middle East. The rate of tall building completions started to rapidly increase in Asia by around the start of 1990’s. By the mid-1990’s similar increases occurred in Europe and the Middle East. In North America tall building completions has been much more cyclical but certainly not insubstantial. By summing the annual values in Figure 5 then there have been 2,729 tall building completions in Europe, 4,996 in Asia, 871 in the Middle East and 4,985 in North America between 1960 and 2017. Some care needs to be taken when interpreting this data as it is likely that some of these buildings have since been demolished. For example, within the database it is known that the 31 storey Red Road flats (ID 42) had been demolished in 2015. Furthermore, as discussed in Section 2.4, around 33 % of incidents include in the database are in buildings that do not meet the 14-storey definition of a tall building suggested by the CTBUH.

11

450 Asia 400 Europe Middle East 350 North America 300

250

200

150

100

50

Approximate number of buildings over 150 150 mcompletedover ofbuildingsnumber Approximate 0

1960 1966 1968 1970 1972 1978 1980 1982 1984 1990 1992 1994 1996 2000 2002 2004 2006 2012 2014 2016 1962 1964 1974 1976 1986 1988 1998 2008 2010 Year

Figure 5 – Approximate number of buildings over 150 m completed between 1960 and 2017, data taken from the CTBUH Tall Building Database.

The second reason for the relatively few incidents in buildings completed between the mid-1970s and the mid-2000s is likely because it is not just the year of construction that is important but whether a building had been refurbished at some point after construction. Within the database there are 13 buildings that have been identified as having been refurbished. A search of the CTBUH façade retrofit database did not reveal any information on the 13 buildings known to have been refurbished. In the case of the Garnock Court (ID 25) incident the exact date of refurbishment has not been determined although it had been “in the last few years” prior to the incident so an estimated year of 1995 has been applied. Figure 6 shows bars for 11 of the buildings where the start is the year of completion and the end is where that building had gone through a refurbishment. The colours of the bars show that all but one of cases have occurred in Europe and that the duration between completion and refurbishment ranges from 14 to 42 years (around 32 years on average) with the refurbishment being applied to buildings completed in the 1960s up to the mid-1970s. The two buildings not included in Figure 6 are both in (ID 14 and ID 15) where refurbishment took place in 2012 however their year of construction has yet to be determined.

12

Figure 6 – Duration between year of building completion and year of refurbishment.

If Figure 4 is updated with (a) the year in which a building was completed, or (b) if it was refurbished then the year of refurbishment, or (c) if the building was still under construction then the year of the incident, Figure 7 results. Here several of the European incidents have moved to the left of figure and additional data are added due to those buildings under construction at the time of the incident, i.e. 5 of the 59 incidents (8 %).

Figure 7 – Year of building completion, refurbishment or construction as a function of number of storeys and region.

Even this is not the full picture as some of the incidents in the database were considerably more significant than others. Some incidents led to multiple casualties and/or significant damage to the building and/or neighbouring property. Thus, the report will return to Figure 7 in the following sections. 2.6 Façade construction In comparison to some of the other information regarding each incident, details on the façade construction were often much more difficult to determine. This is obviously problematic when the main objective of the work is to create a database of façade fires in high-rise buildings. A high level of detail has been identified in very few incidents, whereas some incidents only have a general description of the façade and then in other incidents the level of detail was sparse to non-existent. An 13

additional complicating factor is that it can be difficult to determine materials on simple visual inspection. Information provided within a report may be erroneous and there are known cases of product substitution, i.e. data on what product was installed on buildings maybe uncertain. Thus, the primary database has a descriptive field that is used to provide textual information about the façade. This information may include: the type of assembly, likely materials, and comments regarding the level of confidence in the information sighted. An example of a sparse level of detail is for the 2015 Zikh highway incident in (ID 15) in which the façade is described in one report only as being made of a ‘low quality inflammable material’. On the other hand, the 2008 MGM Monte Carlo Hotel incident (ID 8) has a detailed description of the façade that is given in the database as: “EIFS consisted of a layer of expanded polystyrene foam adhered to gypsum sheathing. Exterior the panels consisted of successive layers of fiberglass mesh and an outside coating of a weather-resistive polymer and cement mixture. EIFS areas had a non-complying thickness of exterior encapsulant. Decorative non-EIFS architectural details of EPS foam encapsulated in polyurethane resin also installed on the exterior”. In a limited number of incidents, specific components and/or manufacturers are named in the reports obtained. It has been decided not to give specific component and manufacturer names in the database as it was not generally possible to verify this information was reliable. Although some of the information is in the public domain identifying specific names may have legal implications and by not providing this within the primary database does not materially affect the outcome of this study.

MCM assembly EIFS assembly Figure 8 – Relationship between face, core, insulation materials and cavity for MCM and EIFS assemblies used for classifying façades within the database.

For the machine learning environment consideration has been made as to how to categorise the façades of the buildings involved in the incidents. It has therefore been decided that five parameters be used to categorise the façade assemblies. Firstly, a façade assembly is classed as either MCM, EIFS or Other if it was unknown. Then four material-related categories have been defined that apply the methodology previously discussed of assessing the façades in terms of their characteristics rather than 14

the components. The four categories are the presence of a ventilated cavity, a face material, a core material and an insulation material. Depending on the specific façade assembly not all four categories will be used. For example, an EIFS may only consist of a face and an insulation material whereas an insulated MCM system will have face, core and also possibly an insulation material. The general relationship between the four material-related categories are illustrated in Figure 8. Examination of the database reveals that 15 incidents (25 %) involved some form of EIFS, 19 incidents (32 %) had some form of MCM assembly and 8 % some other type of façade assembly. Figure 9 shows the breakdown of façade assembly information in terms of the EIFS assembly insulation materials and the MCM panel components. The sparsity of data prevents any further meaningful presentation of the façade material information. EIFS: Combustible 2% EIFS: EPS / XPS 14%

Unknown EIFS: Foam 34% 5%

EIFS: Mineral wool 3% EIFS: PU 2%

MCM: Al / PE Other 24% 8% MCM: Other unknown 8% Figure 9 – Façade assembly types broken-down by insulation type for EIFS and ACM or other components for MCM. It has been decided that including geometrical details regarding material thicknesses, gap widths, etc., not be included in the machine learning environment partly because very few of the incident details give this level of information. 2.7 Cause, origin of fire and age of building when incident occurred The database includes a short description of the cause of the fire for just under 50 % of the incidents and also on which floor the fire was thought to have started for 81 % of the incidents. Both of these elements are not always consistently reported. This may be because of the stage of the incident being reported (i.e. early media reports may not have been precise), difficulty in identifying the cause and/or point of origin during a later investigation, or because of the way in which the floor count is defined, as discussed above. The incident date refers to the day the fire started following the Gregorian calendar in the ISO 8601 [23] format (YYYY-MM-DD). The age of the building at the time of the incident is taken to be the number of years between the construction or subsequent refurbishment of the building and the year 15

that the incident occurred. It is possible that the refurbishment of a building had no impact on the façade but it is clear that in other incidents this was a major factor. Where a building was under construction at the time of the incident the age of the building is taken to be zero years. There are a few incidents where only an approximate date is known, those being BRE Cases 2 and 3 (ID 43 and ID 44), a five-storey residential building in Switzerland (ID 23) and the Butler House incident (ID 33). Four fires have also been identified on the Tamworth Estate (ID 51) where it can only be surmised that these occurred sometime a few years after the buildings were re-clad in the mid-1990s since they were reported in a 1999 document. An analysis of the floor on which the fire started shows that the database has 11 (19 %) unknown incidents, 18 (31 %) that started at the ground level, 5 (8 %) at the roof/top level and the remaining 25 (41 %) at some intermediate floor level. Thus, almost the same percentage of incidents started at ground level and the roof as those that started at some intermediate storey. Figure 10 shows the storey level where the fire was thought to have started against the number of storeys in the building. Fires that started at ground level or at the roof/top level are separately identified, and the maximum possible storey at which a fire could have started is also shown. What is interesting is that very few fires started above half way up the building, although the small dataset does mean it would be prudent not to draw too many conclusions from this.

60

50

40

30

20

Storey Storey where fire started 10

0 0 10 20 30 40 50 60 70 80 90 Number of storeys in the building

Figure 10 – Storey level where the fire was thought to have started against the number of storeys in the building; solid black symbols indicate ground level, solid grey symbols indicate roof/top storey, maximum possible storey where a fire could have started and half way storey height shown by the dashed lines.

For the machine learning environment, the cause has been broken down into seven categories. The categories are based on a subset of the codes used by the UK incident recording system [22] for the source of ignition. The numbers associated with each category description correspond to those used in the UK system. One of the categories is the solid fuel related code used to represent ‘Rubbish’. Although it is recognised that the rubbish would not have ignited itself (unless it self-heated) but would need an actual source of ignition, several incidents are only described in terms of rubbish being the first burning item and so this has been included for the cause of ignition field. The database could be extended in future to include additional categories using the approximately 80 codes defined in the incident recording system.

16

Table 2 and Figure 11 show the number and percentage of incidents associated with each cause category in which electrical equipment makes up almost 20 % of the known causes. There are no suggestions from any of the reports that the cause was deliberate although in some cases the fire was as the indirect result of a deliberate action. For example, the Al Baker Tower 4, (ID 35) incident was as the result of a lit cigarette being thrown off the balcony of the building.

Table 2 – Cause of fire incident machine learning environment categories.

Category (number relates Description Number (%) to equivalent UK incident recording system code)

Unknown_0 Source of ignition unknown 29 (50 %)

Electric_3 Electrical equipment including short circuits 11 (19 %)

Smoking_related_45-46 Smoking materials including cigarette butt 5 (9 %)

Welding_40 Welding activity 3 (5 %)

Fireworks_49 Fireworks 2 (3 %)

Fuel_related_56 Mixed solid fuel items found in rubbish 5 (9 %)

Cooking_appliance_1-7 Cooking appliance including barbecue grill 1 (2 %)

Natural_occurance_52 Lightning 1 (2 %)

Unknown Electric 9% Smoking related 5% 3% Welding 51% 49% 19% 8% Fireworks Fuel related 2% 3% Cooking appliance Natural occurance

Figure 11 – Cause of fire incident machine learning environment categories.

2.8 Casualties Casualties have been defined in terms of fatalities and injuries. Identifying fatality numbers from reports has been relatively unambiguous. However, numbers are subject to some variation depending on when a fatality is attributed to an incident. For example, the official number of fatalities in the Grenfell Tower incident (ID 38) is 71 including a stillborn baby but a 72nd person subsequently died several months later due to complications resulting from the fire.

17

When it comes to injuries it is more challenging since the definition of injury is not so clear. Reports mention that people may have been assessed at the scene with no further action needed. In other cases, people were sent to a hospital for a check-up where either they may have been discharged soon after or some people may have had to stay for a longer period. On examination of several incident reports the number of injuries quoted do not always tally and this is partly attributed to the definition of injury as well as the time of publication of the report relative to the incident. Therefore, any injury numbers in the database should be taken to be indicative values rather than being a precise count. Following on from the numbers of fatalities and injuries there is the question on whether those casualties who were noted as being from the emergency services should be combined with civilians or noted separately. In some incidents it is unclear whether casualties are from one group or another whereas in other incidents the numbers are clearly separated. Thus, for the primary database the extent of any casualties is recorded in a descriptive text field.

Figure 12 – Incidents that involved fatalities and/or injuries.

For the machine learning environment, the numbers of fatalities and injuries are set as numerical fields with values being a combination of those from the emergency services and civilians. Figure 12 shows these data for the incidents identified, with the number of fatalities given in descending order with the associated number of injuries. Incidents where only injuries occurred are given on the right, again in descending order. Fires with fatalities make up 10 of the 59 incidents (17 %) and those with fatalities and/or injuries make up 22 of the 59 incidents (37 %). The analysis in Section 2.5 has examined the number of incidents in terms of the year of completion, year of refurbishment and/or the year a building was under construction. However, as noted previously, the database contains incidents that range in severity and so it could be argued that some 18

of those incidents were cases where the façade contributed very little to the incident. Therefore, Figure 7 has been modified to show those incidents in which fatalities and/or injuries occurred.

90 80 Asia 70 Europe 60 Middle East 50 North America 40 Oceania 30 20

10 Number Number of storeysabove ground 0 1960 1970 1980 1990 2000 2010 2020 Year under construction, year of refurbishment or year completed

Figure 13 – Façade fire incidents by year under construction, year of refurbishment or year completed as a function of building height and region with incidents that incurred fatalities shown with a surrounding box and those that incurred injuries with an x overlaid.

Thus, Figure 13 indicates those incidents where fatalities occurred with a box around the data point and those where injuries occurred are shown with an ‘X’ symbol overlaid. It is clear that those incidents in which fatalities occurred are in buildings with no more than 31 storeys (generally around 100 m). The tallest of the buildings is the Television Cultural Centre, Mandarin Oriental Hotel, Beijing (ID 12), in which a single fire-fighter fatality was reported. This building had 31 stories and a height of 159 m. Furthermore, six of the ten incidents (60 %) were in buildings less than 14 stories and so do not meet the CTBUH definition of a tall building. However, when it comes to injuries there is not a clear pattern to the data. Injuries occurred in the tallest building as well as those that are at the lower end of the height scale. It is important though to reiterate that the definition and identification of injury data was subject to considerable interpretation when entering information into the database. Of the 22 incidents where casualties occurred, eight (~36 %) were in buildings with less than 14 stories. Further investigation of those incidents in which fatalities occurred in respect to the type of façade assembly shows that 6 of 10 has EIFS, 2 had MCM and 2 are classified as Other. In the Daebong Green Apartments (ID 55) incident in which there were four fatalities the façade was described as a "flammable 'styrofoam' material" which may suggest an EIFS assembly. In the Garnock Court, Irvine (ID 25) incident the façade was reported to have exterior walls covered with glass reinforced polyester plastic sheet. If the incidents in which both fatalities and/or injuries occurred are grouped together then the results are shown in Figure 14.

19

Unknown MCM (6) (6) 27% 27%

Other (2) 9%

EIFS (8) 37%

Figure 14 – Number and percentage of façade type assemblies in which fatalities and/or injuries occurred.

It is also important to note that five out of the ten incidents in which fatalities occurred were in buildings that, as already identified in Section 2.5, had undergone refurbishment at a known (or estimated) date after their known date of completion. These incidents were La Tour Mermoz (ID 6), Középszer Street (ID 19), Sonacotra building (ID 21), Garnock Court (ID 25) and Grenfell Tower (ID 38). In addition, the Baku, Binaqadi district building (ID 14) had been refurbished in 2012, although its year of completion has not been determined at the time of writing. Thus, a total of six in ten (60 %) of the fatality fires were in buildings that had been refurbished.

2.9 Sprinklers The primary database gives a description of the performance of any sprinkler system if there was one present during the incident. As the data fields were populated it became clear that most of the incidents identified in the data set were in buildings in which the presence of a sprinkler system was not explicitly noted which suggests none was present. In a small number of incidents, a sprinkler system was present but not always operational and so these cases have been specifically identified. Overall there were 14 incidents in which it appears there was likely to have been a sprinkler system present, six in which the building has been identified as not having a sprinkler system installed and the remaining incidents do not have sufficient information and so are listed as 'Unknown' in the primary database. Specific incidents illustrate the range of outcomes that can occur in the event of a façade fire. The benefit of a sprinkler system mitigating internal fire spread should flames enter the building through broken windows or gaps was demonstrated in the 2014 Lacrosse Tower incident (ID 34). However, in the 2013 -City Towers incident (ID 7) the sprinkler system failed due to the lack of water supply. In the 1997 Eldorado Hotel fire in Nevada (ID 10) the internal sprinkler systems had no effect on the fire since it was limited to the building façade. For the machine learning environment, the presence and activation of sprinklers in the incidents is categorised using a combination of the codes in the UK incident recording system [22] and other relevant fields as shown in Table 3 and Figure 15. It is assumed that no sprinkler system was present in the incident where it has been classified as ‘Unknown’ in the primary database. In five (9 %) of the incidents it was noted that a sprinkler system was present although in some cases it did not activate for some reason. In another eight (14 %) of the incidents the sprinkler system was recorded as having activated but the impact on the fire is not evaluated in the available reports. Two additional categories 20

have been included in Table 3 for any future expansion of the database although none of the incidents were assigned to them at present.

Table 3 – Sprinkler system machine learning environment categories.

Category Description Number (%)

No_sprk_system No sprinkler system installed 45 (76 %)

Sprk_Not_activated Sprinkler system installed but did not activate 5 (8 %)

Sprk_Extingushed_1 Sprinkler system extinguished the fire 0 (0 %)

Sprk_Contained_2 Sprinkler system contained the fire 1 (2 %)

Sprk_Not_contained_3 Sprinkler system activated but did not contain 0 (0 %) the fire

Sprk_Not_known_0 Sprinkler system activated but its effect on the 8 (14 %) fire is unknown

No sprinkler system 14% Did not activate 76% 24% 2% Contained fire 8% Effect unknown

Figure 15 – Sprinkler system machine learning environment categories.

2.10 Wind In this work wind is treated as a natural phenomenon and not in terms of being human created, e.g. specifically where in the Daebong Green Apartments, Uijeongbu incident (ID 55) reports suggest that the down-draught of fire service helicopters contributed to the fire spreading to a neighbouring property. Several wind-related effects are identified from the data. There is the contribution to increased fire growth through fanning the flames, directing flames towards neighbouring buildings or increasing the likelihood of burning pieces of material being dislodged that may aid in the spread of fire. As such, the effect of a strong wind resulting in rapid fire is encoded in the UK incident recording system [22]. However, it is noted elsewhere that the wind in some incidents likely reduced the impact of the fire by moving flames and/or smoke away from buildings. Finally, several incident reports highlight the adverse effect of the wind on fire-fighting operations. For example, in the 2015 Torch Tower incident (ID 2) the wind fanned flames and carried debris to surrounding streets whereas in the 2007 Water

21

Club Tower incident (ID 18) it was reported that the severity of the fire reduced as the wind prevented debris, smoke, and heat from penetrating the building’s interior. The primary database thus uses a text field for the effect of wind in which the information indicates whether wind was not a factor, or how the wind influenced the incident, or if the effects were unknown. The data gives 9 incidents (15 %) in which wind was not a factor, 15 (25 %) in which wind was an influence and 35 (59 %) incidents in which it is not known whether wind was of sufficient significance. For the machine learning environment, the effect of wind is categorised as shown in Table 4 and Figure 16.

Table 4 – Effect of wind machine learning environment categories.

Category Description Number (%)

Increase_fire_growth Caused fire growth to increase due to fanning of 9 (15 %) flames, etc.

Decrease_fire_growth Caused fire growth to decrease due to directing 4 (7 %) flames away from the façade, etc.

Firefighting_operations Major effect of the fire-fighting operations during 2 (3 %) the incident

No_wind_effect Had no or an unknown effect on the fire 44 (75 %) development.

Unknown / no wind effect 7% Increased fire growth 75% 25% 15% 3% Decreased fire growth Impact on operations

Figure 16 – Effect of wind machine learning environment categories.

In four of the ten incidents in which fatalities occurred the database indicates wind had an effect on the fire (two increased the growth due to strong wind, one had an impact on fire-fighting operations, and in one case the wind directed flames away from the building). If injuries are included then an additional two incidents had the fire growth increased by the wind according to observations. 2.11 Falling debris The primary database includes a text field in which any observations regarding falling debris is recorded. As such, several incident reports highlight the adverse effect of falling debris on fire-fighting operations and on the evacuation of people; the increase in fire growth due to burning material or where debris caused property damage. However, in some incidents although falling debris was observed there is no information on whether this had any impact on the fire. The falling debris effects are captured in the machine learning data set through the five categories shown in Table 5 and Figure 17.

22

Table 5 – Effect of falling debris machine learning environment categories.

Category Description Number (%)

Increase_fire_growth Caused fire growth to increase. 8 (14 %)

Caused_property_damage Damage to property, e.g. neighbouring vehicles 5 (8 %)

Firefighting_operations Wind had a major effect of the fire-fighting 4 (7 %) operations during the incident

Observed Falling debris was observed but no specific 8 (14 %) effects identified.

No None or an unknown 34 (58 %)

No falling material Increase 7% 8% fire growth Caused property 58% 42% 14% damage 13% Impact on firefighting operations Observed without specific impact

Figure 17 – Effect of falling debris machine learning environment categories.

All four incidents in which falling debris impacted on fire-fighting operations resulted in fatalities and/or injuries and also four of the eight incidents where fire growth increased also resulted in casualties. 2.12 Analysis There are many over-arching conclusions that can be drawn from the analysis of the data presented in the previous sections. For many of the incidents in which older buildings were involved in a façade fire it is not the year of completion that is important but the fact that several years later these buildings underwent some form of refurbishment. As such, the data shows that 60 % of the fatality fire incidents occurred in refurbished buildings. In many cases, information points towards that refurbishment including some change to the façade design, often the addition of a new cladding or curtain wall assembly which later became a major contributor to the incident. It is less clear whether it was simply the façade assembly components that were the reason the fire developed the way it did or whether it was due to other elements in the façade system, such as the installation of fire stopping, etc. However, it is also possible

23

that the reason for the fire development after refurbishment may have been unrelated to the façade, for example the internal fire separations may have been altered. Considering the number of tall buildings completed in North America since the 1960’s compared to the rest of the world, the number of façade incidents is proportionally less. It is difficult to pin-point the precise reason for this as there may be one or more interacting factors that could be involved. Compared to other regions North America is generally being seen to have a more prescriptive regulatory building code environment, more stringent requirements on professionals involved in the building design process through approvals by Professional Engineers, and the likely higher prevalence of sprinkler system installation. However further study would be required to assess whether the frequency of fires in high-rise building differs to other regions, whether the types of façade assemblies differ, etc. When a focus is placed on those incidents in which casualties occurred then a number of interesting findings arise. In the incidents in which fatalities occurred the buildings were generally less than 30 storeys tall and had an EIFS-type assembly façade. This is probably somewhat surprising as much of the recent media focus has been on the incidents in which ACM assemblies have been involved. However, the incident in which the most fatalities occurred was at Grenfell Tower (ID 38) in which the façade system did consist of an ACM assembly. If injuries are also included then the range of heights of the building varies further than when just fatalities are considered but EIFS-type assemblies are still the most predominant, although the façade assembly is not known for 27 % of the incidents. What is also noticeable in those incidents in which casualties occurred is the impact of wind and of falling material. Both factors appear to have contributed to the outcome through enhancing the development of the fire and impacting on fire and rescue operations. It is difficult make major conclusions from the limited number of incidents in which sprinkler systems were installed in buildings and their effect on façade fire incidents. Of 14 incidents with a sprinkler system installed it was found that in five cases there was no activation. However, of those fire cases it is reasonable to suggest that in the Eldorado Hotel, Nevada incident (ID 10) the system was not expected to activate since the fire was limited to the external part of the building. In the Wooshin Golden Suites, incident (ID 11) it was reported that although the building had a sprinkler system there were no sprinklers in the room of fire origin. Thus, sprinkler systems could be said to have failed as might be intended in three of the 14 incidents, i.e. suggesting a reliability of around 80 % which is within the range found by Frank et al. [18].

24

3 MACHINE LEARNING Machine learning is an emerging field which provides an expanded toolbox with tools such as neural networks and support vector machines to systematically mine complex datasets for patterns and structures. There are a number of programming libraries and software packages available including TensorFlow, Microsoft Cognitive Toolkit and scikit-learn. These tools can potentially be used to apply machine learning methods to a façade fire database to predict some form of ‘severity’ for a given high- rise building if a fire incident should occur. Examples of how machine learning and the types of algorithms that are transforming knowledge generation in other fields are shown in Table 6.

Table 6 – Different machine learning algorithm types and their application.

ML model type Algorithm Application

Supervised Regression Advertising popularity prediction, weather forecasting, learning market forecasting, estimating life expectancy

Classification Identity fraud detection, image classification, real-time graphics detection

Unsupervised Dimensionality Big data visualisation, meaningful compression, learning reduction structure discovery, feature elicitation

Clustering Recommender systems, targeted marketing, customer segmentation

Reinforcement - Real-time decisions, robot navigation, learning tasks, learning skill acquisition, game artificial intelligence

Following the above-mentioned machine learning tools, TensorFlow is used in this study with the following advantages: • TensorFlow can be used for supervised learning. The available data of this study comprises multiple feature data with labels (i.e. labelled data) which fits the realm of supervised learning; • TensorFlow is available as a Python library. Python is a well-known programming language for scientific analysis, its open-source and cross-platform nature means the developed model will be executable on multiple operating systems (i.e. Linux, Windows and Mac OS). A wide range of available Python libraries (i.e. NumPy and SciPy) provide extended possibilities of data manipulation options using TensorFlow; • Similar to several other tools, TensorFlow is free and open source. Its source code and compiled packages are widely available on a website providing future development opportunities for others without additional cost; and • Documentation and an extensive user-based community are available for TensorFlow. Quality documentation enhances the learning process for researchers.

3.1 Software specification The currently published version of TensorFlow supports only 64-bit machines with advantages of employing high RAM capacity. A 32-bit version of TensorFlow can be created by building directly from its source code. The models in this study are built on the Windows 10 64-bit operating system through

25

the use of 64-bit TensorFlow and 64-bit Python. Table 7 shows the application/library versions that are used for building the machine learning models demonstrated in this study. The indicated software specifications are prerequisites for running the source code associated with this report.

Table 7 – Application specification used for this study.

Software Version

Python 3 (64-bit) 3.6.4

TensorFlow 1.7.0

TensorBoard 1.7.0

NumPy 1.14.2 pandas 0.22.0

3.2 TensorFlow application The current usage of TensorFlow includes image classification and speech recognition [24], computational biology [25] and many others. There are several published TensorFlow model source codes which share similarity to the problem being addressed in this study, one of them is by Google and documented as an official tutorial [26]. One application of machine learning would be to allow the assessment of new or existing façade systems and estimate, if a fire occurred, what would the likelihood be that it could become a major incident. This approach would need the database to have a ‘pass’ or ‘fail’ vectors where each incident is assigned one of the two options. However, as it stands with the primary database only having predominantly major incidents means there are few ‘pass’ data sets. What the database does record is the number of casualties that occurred during incidents and whether those casualties were injuries or deaths. Thus, a machine learning model can be trained to exploit the correlation between the Severity (i.e. measured as resulted fatalities and injuries) and a given building with a specified façade.

Figure 18 – The TensorFlow programming environment. [tensorflow.org]

Figure 18 illustrates the TensorFlow API hierarchy, from the base level TensorFlow engine to the high- level ‘user-friendly’ estimators. The TensorFlow engine provides essential routines which handle the fundamental TensorFlow tasks and currently they can be accessed via Python, C++, Java and Go. The 26

functions and classes that TensorFlow provides in its Python binding are grouped into three layers: low, mid and high-level APIs. The low-level TensorFlow APIs provide the most versatile way for building a model but often requires more effort in doing so. The high-level TensorFlow APIs provide some user- friendly packaged modules allowing users to time efficiently build a model, with limited customisation options. 3.3 Classification model trained on the surveyed dataset 3.3.1 Overview A TensorFlow classifier model consists of an input layer that supplies data as ‘features’ as an input layer to a series of hidden layers each with a defined number of neurons. These hidden layers and neurons are interconnected and then connected to ‘labels’ as an output layer that provides the predictive results from the model. All the hidden layers and neurons can be interconnected or some connections can be omitted. TensorFlow includes several estimator classes that can be used to carry out different types model fitting algorithms and Table 8 shows the available classes in the Estimator module.

Table 8 – Key model classes contained in the TensorFlow Estimator module.

TensorFlow Estimator classes Description

LinearClassifier An estimator for TensorFlow linear classifier model

LinearRegressor An estimator for TensorFlow linear regression problems

DNNClassifier A classifier for TensorFlow deep neural network (DNN) models DNNRegressor A regressor for TensorFlow deep neural network (DNN) models

DNNLinearCombinedClassifier An estimator for TensorFlow linear and deep neural network (DNN) joined classification models

DNNLinearCombinedRegressor An estimator for TensorFlow linear and deep neural network (DNN) joined models for regression

The objective has been to generalise the input parameters of the surveyed data, as demonstrated in Section 2, in-line with provided labels. Given the available data structure, it has been decided that a supervised learning model would be the most suitable option. The DNNClassifier estimator class has been adopted, since this class is capable of modelling both linear and/or non-linear problems by adjusting the number of hidden layers and neurons. The method used to determine the number of hidden layers and neurons is discussed in Section 3.3.5. The surveyed data have been converted to TensorFlow features and label columns before being used for training and testing a model, as further described in Section 3.3.2. For each test incident case, the model returns ‘probabilities’ for each defined class, the class with the highest probability is then selected as a predicted Severity classification. The output layer has a total of 11 Severity classifications where the Severity is derived from the Injury and Fatality data in the survey. The 11 Severity classes in the output layer have been broken into cases where either there were no fatalities but zero or more injuries occurred or where fatalities occurred. Such a method simplifies how Severity is classified,

27

without generating too many classifications but still makes it possible for the model to recognise the data patterns. Section 3.3.3 addresses how the Severity classification is calculated. Models have been trained with different numbers of hidden layers and neurons per hidden layer, details of the selected values for these parameters can be found in Section 3.3.5, to optimise the model parameters. The model training and testing process is outlined as below: 1. Parsing training data and testing data – in this case, a predefined number of training data are selected from the entire machine learning database, what is left is used as testing data. For example, if 40 data entries are randomly selected from a database which contains 60 data entries; the remaining 20 data entries will be used for testing the model after training is complete; 2. Create feature and label columns – for both training and testing to provide data supply pathways to the model, allowing raw data to be structured in a form to fit the model requirements; 3. Model training – the model is fitted for the selected training data; and 4. Model testing – the model accuracy is evaluated based on the selected testing data.

3.3.2 Input data preparation The surveyed data has been ‘cleaned’ before they are used for modelling. This process is primarily to reduce potential human error when putting the data together by hand (i.e. typographical errors) and also to make sure that the input data format is compatible with TensorFlow. The process comprises four steps: 1. All strings are converted to lower case letters to prevent unintended errors. For example, the strings ‘floor’ and ‘Floor’ are recognised as two distinct elements even they are meant to have the same meaning; 2. White spaces in headers are replaced with underscore ‘_’ as the TensorFlow feature object does not allow a name to contain white spaces; 3. In the primary database the word ‘unknown’ is used where the relevant information is uncertain or absent. All data under each feature must be a single data type, thus for example, a text string of ‘unknown’ under building age would render data type inconsistency. Hence the text string ‘unknown’ is replaced with a non-existing number, in this case ‘-1’ is used to replace text ‘unknown’; and 4. Category values in the primary database with ‘n/a’ string used to indicate a field is ‘not applicable’ are replaced with ‘-2’ with the same intention stated above for ‘unknown’.

The surveyed data have been converted to TensorFlow features and label columns before being used for training and testing a model. The original surveyed data have 18 properties for each incident and 12 of these properties have been used for constructing TensorFlow feature objects. The 12 properties are selected based on the fact that they are pre-incident parameters (parameters that can be obtained for a building before fire incident, for example, ignition source is omitted because it is unknown before incident). This allows the trained model to predict the Severity classification for a given existing high- rise building. The neglected data are: building age at the incident, ignition source, floor level of the ignition, implication of wind condition during the fire and observed falling material during the fire. Appendix D shows the specifications of the 14 input feature columns.

28

3.3.3 Features and label Feature columns are the TensorFlow Estimator’s data mapping objects, they are provided to specify the data type of the supplied database and must be specified to allow the program to tackle the input data accordingly. For example, the age and the façade type of a building are defined in the format of numbers and text, respectively. Table 9 shows the three feature types which are used in the model. Training a machine learning model can be regarded as fitting a function based on a set of given input parameters 푥1 to 푥푛 and the related outcome 푦, as given in Equation 1. In machine learning terms, input parameters 푥1 to 푥푛 are features and 푦 is the label. Data provided in such form is labelled data.

푦 = 푓(푥1, 푥2, 푥3 … 푥푛) Equation 1 Table 9 – Feature column types which are used for constructing the model.

Feature column type Description

Numeric column For numeric data type, can be integers or floating-point numbers

Categorical column with a For a defined list of categories, i.e. a column to describe specified vocabulary list façade built type

Bucketised column This can be used alongside the numerical column to group a range of numbers into defined conditions, i.e. a predefined year boundary [1970, 1980, 1990, 2000] will be converted to a Boolean list [0, 1, 0, 0] for a building built in 1981. i.e. [1970 ≤ year < 1980, 1980 ≤ year < 1990, 1990 ≤ year < 2000, 2000 ≤ year].

In this case, the label (Severity of an incident) is not directly supplied in the surveyed data but is calculated based on Fatality and Injury, shown in Table 10.

Table 10 – Data used for calculation of Severity classification (i.e. label column).

Parameter Description Data type

Fatality Total fatality number for the incident int

Injury Total injury number for the incident int

The calculated Severity is shown in Table 11 where the classification method is not solely designed based on the available data but also considers potential future data expansion. As such, the existing data does not generate Severity classifications 5 and 6.

29

Table 11 – Risk classification rules, calculated based upon surveyed fatality and injury.

Severity class (푺) Fatality (풗풇) Injury (풗풊)

0 푣푓 = 0 푣푖 = 0

1 푣푓 = 0 1 ≤ 푣푖 < 5

2 푣푓 = 0 5 ≤ 푣푖 < 10

3 푣푓 = 0 10 ≤ 푣푖 < 20

4 푣푓 = 0 20 ≤ 푣푖 < 50

5 푣푓 = 0 50 ≤ 푣푖

6 1 ≤ 푣푓 < 5 푎푛푦

7 5 ≤ 푣푓 < 10 푎푛푦

8 10 ≤ 푣푓 < 20 푎푛푦

9 20 ≤ 푣푓 < 50 푎푛푦

10 50 ≤ 푣푓 푎푛푦

3.3.4 Machine learning model In order to identify the optimal model, a range of configurations for various numbers of hidden layers, neurons for each hidden layer and training steps have been investigated. All accuracies of these tested model configurations are documented with discussions of these results in Section 3.3.5.

30

Figure 19 – Example of one of the trained model’s architecture of the deep neural network (DNN) model trained on surveyed data.

Figure 19 shows one of the trained model architectures. In this example model, 4 hidden layers and 5 neurons on each hidden layer are adopted. The model comprises a total of 14 features in the Input Layer. Post-incident properties are neglected as the objective of this model is to estimate its Severity level prior to an incident. For each training step, a batch of feature data is supplied to the model via the input layer which then flows through the hidden layers. Then, when data flows through a neuron, the input value is multiplied by a weight which measures the significance of how the preceding neurons influence the current neuron. In Figure 19 for example, each of the arrowed lines is associated with a weight. A bias is applied to the sum of all weighted neuron values. This is to impose a threshold number to control how easily the summed values can be activated by the activation function. Equation 2 shows the calculation of each neuron on a hidden layer such that 푛

푦푗 = 휎 (∑ 푤푖푥푖 + 푏푗) 푖=0 Equation 2 where: 푦 is the resultant value of a given neuron on hidden layers;

푥푖 is the value of the preceding neuron which connected to the given neuron;

푤푖 is the weight of the preceding neuron which connected to the given neuron; 푏 is the bias; and 휎 is an activation function.

31

Results at the ten output neurons are directly obtained from the hidden layers in the range from zero to infinity. Applying a ‘softmax’ function normalises the output neuron values to 0 – 1. This is shown in Figure 19 as the probability/confidences of each Severity classification in the output layer. At the end of each training step, the model adjusts the neuron biases and weights to minimise the total loss by gradient descent. The total loss of the model after each training step is evaluated by following Equation 3. Where the average loss is used, it is calculated by dividing the total loss by the total number of neurons in the output layer (i.e. the total number of Severity classifications). Thus 푛

퐿푡표푡푎푙 = ∑ 푃푝,푖 − 푃푎,푖 푖=0 Equation 3 where: 퐿푡표푡푎푙 is the total loss;

푃푝,푖 is predicted probability of classification 푖;

푃푎,푖 is actual probability of classification 푖; and 푛 is the total number of classification or the total number neurons in the output layer.

30

25

s

s

o

l

g 20

n

i

n

i a

r 15

t

e

g a

r 10

e

v A 5

0 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 Training step [10³]

Figure 20 – Training loss against training step, extracted from one of the trained cases with 2,000 training steps.

Figure 20 shows an example of the averaged training loss of one of the preliminary models (2 hidden layers with 12 and 10 neurons on each layer) as the number of training steps is increased. It is observed that the model is properly tuning the weights and biases achieving lower loss as training progresses. However, the average training loss is assessed based on training datasets. Greater training steps do not necessarily result in higher testing accuracy, the model can be over-fitted to the training data as the number of training steps increases. Therefore, the optimised number of training steps should be sufficient to generalise the data pattern without over-fitting. One of the techniques in machine learning to address this issue of over-fitting is to randomly drop some of the connections (i.e. the arrowed lines in Figure 19) [27]. However, this technique is not applied in this study and the model can be potentially benefit by adopting this feature in future versions.

32

After the model is trained, it can be evaluated based on testing which uses independent datasets separated from the training data. As a final product, the trained and tested model will be capable of making predictions for given inputs in line with the features that the model is trained for (i.e. the features shown in Figure 19). 3.3.5 Selection of optimal model parameters As mentioned previously in Section 3.3.4, a series of models with different configurations are trained to identify the best performing model configuration. The variable model parameters are shown in Table 12. As a result, 256 models with unique configurations are generated and trained.

Table 12 – Tested model parameters for selecting the appropriately optimal model configuration.

Variable Tested value Potential impact on a neural network model

Number of [1, 2, 3, 4] More hidden layers allow the model to identify more hidden layers complex data patterns, i.e. data pattern based on data pattern of previously hidden layers but increases training time

Number of [2, 3, 4, 5] More neurons will allow the model to identify more possible neurons per data patterns (i.e. correlations), but increases training time hidden layer

Training steps [0.5, 1, 2, 4] Influences the fitness of the model to the training data; a (1000’s) greater number of training steps may result in overfitting

Training [6, 18, 30, 42] Proportion of the entire database which are used for training incidents purpose and the remainders are used for testing the trained model

One of the challenges arising from the limited database counts is the ability to properly present the performance of a trained model. Specifically, the testing data is limited by both the total available data and the data used for training. For example, a model has potentially test accuracies of 0, 50 or 100 % if only two incidents are available for testing. To address this issue, a ‘random shuffle and re-evaluate’ process is adopted in which every model is trained and evaluated multiple times, each with unique initial condition and randomly selected training datasets (with the remaining data used for testing). This process is repeated 20 times for each model, this number is selected to keep computing time within a reasonable period, limited by available hardware capabilities. Finally, the performance of each model is evaluated from the averaged performances of the 20 individual performances. Averaged accuracies for all trained models are documented in Appendix C.

33

Table 13 shows the top performing model configurations with averaged accuracy greater than 60 %. The following characteristics for each variable parameter are observed: • Training steps: the minimum training steps tested appears to be best at generalising the data correlations, higher training steps may have resulted in over-fitting in this case.; • Training incidents: a greater number of training data is always preferred and the results indicated the highest tested number of incidents data gives the best performance. The most poorly performing model configurations (accuracy as low as 38 %) are supplied with minimum of six training incidents (approximately 10 % of available incidents). 42 training data (approximately 70 % of available incidents) is, therefore, the best option; • Number of neurons per layer: Two neurons is observed to be the best performance; and • Number of layers: a single layer appears to be adequate for the model.

Table 13 – Selected best performing trained models with testing accuracies and model parameters.

Testing Training steps Training No. of neurons per No. of accuracy incidents layer layers

64 % 500 42 2 1

63 % 500 42 3 3

62 % 1000 42 2 1

61 % 500 42 2 4

61 % 500 42 3 4

61 % 500 42 3 2

61 % 2000 18 2 4

The optimised model parameters are selected based upon observations explained above, shown in Table 14.

Table 14 – Optimised parameters for the DNN model.

Model parameter Optimised value

Training steps 500

Training incidents 42

No. of neurons per hidden layer 2

No. of hidden layers 1

3.3.6 Results of optimised model and observations The model training process is successful in terms of minimising the loss. Figure 20 clearly shows a decreased average loss as the number of training steps increases, implying the model is being fit according to the training dataset. The average loss rate reaches constant after approximately 10,000

34

training steps. However, the model over-fits after 500 to 2,000 training steps depending on model configuration.

Figure 21 – Model predicted against actual Severity classification with surveyed incident index labelled.

Figure 21 shows predicted Severity classification of the best-performing optimised model, evaluated on all surveyed incidents (where the numerical results can be found in Appendix E). There are four incidents (IDs 6, 55, 57 and 38) where the match is poor, several where there is some variation (IDs 2, 9, 40 etc.) and then the rest in which the predictions correlate to the actual incident Severity. Thus, the optimised model is able to predict the Severity classification with 63 % accuracy when evaluated based on test data (i.e. independent from training data). The optimised trained model exceeds the ‘random guess’ accuracy of 11 % (excluding Severity classifications 5 and 6) where for example, by randomly guessing a multiple-choice question consisting of ten options, the guessed answer has 10 % chance to get the correct answer. However, the accuracy is biased in that many of the incidents have an actual Severity classification of zero and the model almost always predicts a zero Severity when the match is poor. Thus, were there are more incidents with a non-zero Severity it seems that the model accuracy would likely decrease. This suggests that in future it would be beneficial to revisit the Severity classification to create one that has less bias towards a single result. One way to do this might be to place severity in terms of a monetary sum such as incorporating damage in the consequence. The trained model and Python source code can be found in the GitHub repository github.com/ PYSFE/ctbuh_alpha.

35

4 CONCLUSIONS AND FUTURE WORK A database of façade fire incidents has been created in this study. Although a number of the incidents identified have already been discussed in the literature, this work has created a structured dataset that allows for some level of numerical analysis to be conducted. Most of the incidents included in the database are those that resulted in some significant outcome whether that be casualties and/or damage to the building and neighbouring property. Such incidents are often covered by media organisations. However, the level of detail and the accuracy of the reports are sometimes lacking. A manual analysis of the database has identified several interesting findings regarding the buildings in which incidents occurred, where casualties have resulted, the types of materials used in façade assemblies and some tentative findings on sprinkler system performance. In summary the findings are: • In the case of incidents involving older buildings it is not the year of completion that is important but that several years later these buildings underwent some form of refurbishment. Findings show that 60 % of the fatality fire incidents occurred in such buildings; • In incidents in which fatalities occurred the buildings were less than 32 storeys tall and had an EIFS-type assembly façade. When injuries are included in the analysis then it is more difficult to identify building height and façade assembly trends in the data; • The number of façade incidents in buildings completed in North America are proportionally less than the rest of the world; • Based on the interpretation of very limited data the reliability of sprinklers systems in façade fire incidents is around 80 %.

The database presents a snapshot of information that the authors have managed to source during the study period. It is likely that further existing information about incidents will come to light with more investigation and new information will be released after the completion of the study. Therefore, it will be valuable to enhance the current database by filling in gaps, where possible. Furthermore, it may be beneficial to consider extending the scope of the database to include more information about other building systems such as automatic fire detection and alarm systems, the passive fire protection in the building, both of which may have a major impact on fires that have the potential to involve casualties. Following on from the creation of the primary database, the TensorFlow machine learning software has been used to demonstrate how such a tool might be used to provide a predictive method that relates how the type of building and façade might result in casualties should a fire occur. The primary database has been used to create a set of input features and a pilot Severity classification has been proposed to assess the predictive capability of the machine learning environment. Various combinations of models have been assessed and the best-performing optimised model gives a 63 % accuracy against the original incident data. However, the limited scope of both the Severity classification and the scarcity of available input data means results should only be indicative of the potential application of the machine learning tool. The trained models are deep neural network classifier type. This type of model has proven to be of high performance for non-linear problem with large and complex data structure. Whilst the available data maintains a relatively complex form (i.e. consists of more than 10 properties), the number of available surveyed incidents is limited. Therefore, other modelling types (i.e. a linear or combined model) may result in improved performance. Currently the trained models are fully packed without inspecting the individual neurons. This is sufficient for the given objective of making predictions with appropriate accuracy but not helpful in understanding how the model derived such a solution. Future investigation of detailed model individual components will enhance the understanding of the importance of the input data parameters.

36

In order to understand why façade fire incidents can develop into major events it would be beneficial to identify those incidents that involved façades which did not significantly develop. This is a challenge since fires that do not become major incidents do not get reported in the media, etc., to the same extent. There also needs to be a judgement made on whether the incident did involve or had the potential to involve the façade to a sufficient degree to be classed as a ‘façade fire’. However, if enough of these incidents could be identified then there is the potential to train the machine learning system to assess the probability of a given façade installation developing into a major fire event. There has been preliminary discussion with a major UK fire and rescue service to see if it is possible to identify these types of incidents but at the time of writing no further progress has been possible. There should be consideration on how the façade fire database can be taken forward. One option would be to make it publicly available by simply creating a wiki-type webpage that can be exported and modified by anyone. Alternatively, the database could be hosted by the CTBUH in a similar way to the façade retrofit database. The high-rise building database could then be linked to the façade fire incident database much like what could be done between the CTBUH tall building database and the CTBUH façade retrofit database. Either way, since the data is currently sparse, based on the information that has been found publicly, then making the database open for other parties to contribute it may be filled out more completely. However, there would need to be some form of review process the validate any changes or additions to the database.

5 ACKNOWLEDGEMENTS The authors would like to thank the CTBUH and Sun Hung Kai Properties for the seed funding that was made available to carry out this study.

6 REFERENCES 1. London fire: What happened at Grenfell Tower? BBC News, 19 July 2017, www.bbc.co.uk/news/uk-england-london-40272168 2. Delichatsios, M.A. Enclosure and facade fires: physics and applications. Fire Safety Science 11: 3- 27, 2014. 10.3801/IAFSS.FSS.11-3 3. Holland C, Crowder D, Shipp M. External fire spread – Part 1 Background research. BRE Global, Garston, April 2016. 4. Holland C, Crowder D, Shipp M, Cole N. External fire spread – Part 2 Experimental research. BRE Global, Garston, April 2016. 5. Moss, A. G. Façade fire spread in multi-storey buildings. BRANZ Study Report 32, Building Research Association of New Zealand, Judgeford, New Zealand, 1990. 6. Wade C. A., Clampett J. Fire performance of exterior claddings. BRANZ Report FCR 1, Building Research Association of New Zealand, Judgeford, New Zealand, 2000. 7. White N., Delichatsios M. Fire hazards of exterior wall assemblies containing combustible components. Fire Protection Research Foundation, 2014. 8. NFPA 285. Standard fire test method for evaluation of fire propagation characteristics of exterior non-load-bearing wall assemblies containing combustible components. National Fire Protection Association, Quincy, MA, USA, 2012. 9. BS 8414-1:2015 + A1:2017. Fire performance of external cladding systems. Test method for non- load bearing external cladding systems applied to the masonry face of a building. British Standards Institution (BSI), London, UK, 2015.

37

10. ISO-13785. Reaction-to-fire tests for façades, International Organization for Standardization, 2002. 11. ANSI/FM 4880-2017. American National Standard for evaluating the fire performance of insulated building panel assemblies and interior finish materials. FM Approvals LLC, Norwood, MA, USA, 2017. 12. Colwell, S., Baker, B. Fire performance of thermal insulation for walls of multi-storey buildings (3rd ed.). BRE 135, BRE, Garston, UK, 2013. 13. Gabel, J. Another record-breaker for completions; 18 “Tallest Titles” bestowed, 2016, www.skyscrapercenter.com/research/CTBUH_ResearchReport_2016YearInReview.pdf 14. Paech, C. Structural membranes used in modern building facades, Procedia Engineering 155, 61 – 70, 2016. 15. Babrauskas, V. Glass breakage in fires, Fire Science and Technology Inc., 2010. www.doctorfire.com/GlassBreak.pdf 16. Wang, Y., Wang, Q., Shao, G., Chen, H., Su, Y., Sun, J., He, L., Liew, K. M. Fracture behavior of a four-point fixed glass curtain wall under fire conditions. Fire Safety Journal, 67, 24-34, 2014. doi.org/10.1016/j.firesaf.2014.05.002. 17. Valiulis, J. Building exterior wall assembly flammability, fireengineering.com, 2015. www.fireengineering.com/content/dam/fe/online-articles/documents/2015/Valiulis.pdf 18. Frank, K. M., Gravestock, N., Spearpoint, M. J., Fleischmann, C. M. A review of sprinkler system effectiveness studies. Fire Science Reviews, 2(6), 2013. 10.1186/2193-0414-2-6 19. An, W., Huang, X., Wang, Q., Zhang, Y., Sun, J., Liew, K. M., Wang, H., Xiao, H. Effects of sample width and inclined angle on flame spread across expanded polystyrene surface in plateau and plain environments, Journal of Thermoplastic Composite Materials, 28(1), 111-127, 2013. 10.1177/0892705713486132 20. Melbourne Fire Brigade, Lacrosse Docklands, Post incident analysis, Report 1403134A, undated. 21. Evans, D. High-rise façade fires: a world-wide concern. Fire Safety in Towers Conference, Sept 2017. 22. Incident Recording System – Questions and Lists, Version 1.6 – (XML Schemas v1-0p), Department for Communities and Local Government, London, 2012. 23. ISO 8601, Data elements and interchange formats – Information interchange – Representation of dates and times. 24. Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., Devin, M., Ghemawat, S., Irving, G., Isard, M., Kudlur, M. TensorFlow: A system for large-scale machine learning. OSDI, 16, 265-283, 2016. 25. Rampasek, L., Goldenberg, A. Tensorflow: Biology’s gateway to deep learning? Cell systems. 2(1), 12-4, 2016. 26. Cheng, H. T., Koc, L., Harmsen, J., Shaked, T., Chandra, T., Aradhye, H., Anderson, G., Corrado, G., Chai, W., Ispir, M., Anil, R. Wide & deep learning for recommender systems. Proceedings of the 1st Workshop on Deep Learning for Recommender Systems, ACM, 7-10, 2016. 27. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15, 1929- 1958, 2014.

38

APPENDIX A : TERMINOLOGY The complexity of modern buildings means that the terminology associated with façades used in the various documents referenced in this report are not wholly consistent with one another. The following terminology has been applied in this report.

• Material – a substance from which an item is or can be made, for example polystyrene.

• Composite – two or materials bonded or mixed together, for example metal composite material (MCM).

• Panel – a flat or curved item, typically rectangular made of a material or composite, for example a MCM panel.

• Render – a coat (or layer) of a continuous material.

• Component – general term for a type of panel, a type of render, etc.

• Assembly – a combination of components including fixings, panels and/or render, moisture barrier, paint etc.

• Curtain wall – an assembly suspended from the building structure.

• Cladding – assembly attached to building structure.

• Façade system – an assembly or assemblies plus balconies, windows etc.

• Façade – a general term to describe the exterior of a building.

39

APPENDIX B : INCIDENT DATA

40

Name The Address, Dubai ID 1 Coordinates (Longitude, latitude, elevation above sea level) (25.1936328, 55.2790710, 3) Incident description The building contains a mixture of residential and hotel accomodation. The fire was started by an electrical short-circuit. It was reported that 14 people suffered minor injuries, one person was moderately injured and another had a heart attack due to overcrowding and smoke at the site. A strong wind blew some of the fiercest flames away from the building, but burning panels drifted up to a street block away and ignited secondary fires on adjacent roofs, despite the fire service’s rapid attendance and their hosing down rooftops from tall aerial platforms. Investigations suggest that the facade panels were tested in 2007 to ASTM E-119 in combination with gypsum board which provided the fire resistance rather than being tested to NFPA 285. Primary database content Machine learning data Date 2016-01-01 Age 8 Floors 63 Floors 63 Type Residential / hotel Type Hotel Height (m) 302 Height 302 Country UAE Country UAE Year built 2008 Year built 2008 Year refurbished n/a Under cons. No Construction type Concrete Construction Concrete Façade description Type MCM Cladding described as aluminium plastic composite panel with 'anti-toxic' Cavity Unknown polyethylene core material. Face Aluminium material Core PE material Insulation Unknown material Cause Short circuit was caused by electrical wires of the spotlight used to Cause Electric_3 illuminate the building between the 14th and the 15th floor

Starting floor 20 Starting floor 20 Casulties 16 injuries Fatalaties 0 Injuries 16 Sprinklers Sprinkler Sprk_Not_known_0 A sprinkler system was installed but there were conflicting reports on whether it activated or not. Some reports say the activated system ran out of water after 15 min. Wind as a factor Yes Wind Decrease_fire_growth

Falling material as a factor Falling material was reported to have caused the Falling material Increase_fire_growth fire in a lower flat and threatened people below the building. Sources 1. Evans D. High-rise façade fires: a world-wide concern. Fire Safety in Towers Conference, Sept 2017. 2. www.telegraph.co.uk/news/worldnews/middleeast/dubai/12076792/Dubai-skyscraper-fire-new- years-eve-2015-live.html 3. www.dailymail.co.uk/news/article-3380507/Fire-rips-Dubai-hotel-close-city-hold-New-Year- celebrations.html 4. www.thenational.ae/uae/dubai-hotel-s-sprinklers-ran-out-of-water-15-minutes-into-fire- 1.207750 5. Fire risks from external cladding panels – A perspective from the UK, Probyn Miers Ltd, 2018.

6. www.skyscrapercenter.com/building/the-address/468

41

Name The Torch, Dubai ID 2 Coordinates (Longitude, latitude, elevation above sea level) (25.0879420, 55.1475000, 1) Incident description Reports imply that the building was clad with ACM. Fire started around the 50th floor on one of the building’s balconies and burned until it ultimately reached the roof. The fire was fanned by high winds and burning panels were seen falling from the building. 101 apartments were damaged by the fire.

Primary database content Machine learning data Date 2015-02-21 Age 4 Floors 86 Floors 86 Type Residential Type Residential Height (m) 352 Height 352 Country UAE Country UAE Year built 2011 Year built 2011 Year refurbished n/a Under cons. No Construction type Concrete Construction Concrete Façade description Type MCM Reports imply that the building was clad with ACM. Cavity Unknown

Face Aluminium material Core PE material Insulation Unknown material Cause Barbecue grill on 51st floor balcony Cause Cooking_appliance_1-7

Starting floor 51 Starting floor 51 Casulties 4 injuries Fatalaties 0 Injuries 4 Sprinklers Sprinkler Sprk_Not_known_0 Probyn Miers report suggests that sprinklers were present but activation was 'delayed'.

Wind as a factor Wind fanned flames and carried debris to Wind Increase_fire_growth surrounding streets Falling material as a factor Videos show falling burning material starting a Falling material Increase_fire_growth secondary fire at a lower level.

Sources 1. Evans D. High-rise façade fires: a world-wide concern. Fire Safety in Towers Conference, Sept 2017. 2. Fire risks from external cladding panels – A perspective from the UK, Probyn Miers, 2016. www.probyn-miers.com/perspective/2016/02/fire-risks-from-external-cladding-panels- 3. www.stuff.co.nz/world/middle-east/66484552/79storey-dubai-tower-engulfed-in-flames

4. www.dailymail.co.uk/news/article-2962507/Towering-inferno-Thousands-evacuated-fire-rips- one-world-s-tallest-residential-buildings-Dubai.html 5. Fire risks from external cladding panels – A perspective from the UK, Probyn Miers Ltd, 2018.

6. www.skyscrapercenter.com/building/the-torch/344

42

Name The Torch, Dubai ID 3 Coordinates (Longitude, latitude, elevation above sea level) (25.0879420, 55.1475000, 1) Incident description The second fire in the building since the previous one in 2015. The exterior cladding was reportedly being renovated at the time of this incident. The fire rapidly reached the roof and an entire side of the building became engulfed in flames.

Primary database content Machine learning data Date 2017-08-04 Age 6 Floors 86 Floors 86 Type Residential Type Residential Height (m) 352 Height 352 Country UAE Country UAE Year built 2011 Year built 2011 Year refurbished Ongoing since previous incident Under cons. No Construction type Concrete Construction Concrete Façade description Type MCM Reports imply that the building was clad with ACM. Cavity Unknown

Face Aluminium material Core PE material Insulation Unknown material Cause Unknown Cause Unknown_0

Starting floor Unknown Starting floor Unknown Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler Sprk_Not_known_0 Reports on previous incident suggest sprinklers were present

Wind as a factor No Wind No_wind_effect

Falling material as a factor Debris seen falling to the ground. Falling material Observed

Sources 1. Evans D. High-rise façade fires: a world-wide concern. Fire Safety in Towers Conference, Sept 2017. 2. Fire risks from external cladding panels – A perspective from the UK, Probyn Miers, 2016. www.probyn-miers.com/perspective/2016/02/fire-risks-from-external-cladding-panels- 3. www.telegraph.co.uk/news/2017/08/03/dubai-skyscraper-fire-1100ft-torch-tower-engulfed- flames/ 4. www.nytimes.com/2017/08/03/world/middleeast/torch-tower-dubai-fire.html

5. www.skyscrapercenter.com/building/the-torch/344

43

Name Tamweel Tower, Dubai ID 4 Coordinates (Longitude, latitude, elevation above sea level) (25.0797690, 55.1532920, 0) Incident description The Tamweel Tower is a mixed use and residential building. The fire was thought to have started by a cigarette discarded onto pile of waste materials left by workers. It spread into the building damaging 9 floors and 9 apartments and also to some cars parked below due to falling debris.

Primary database content Machine learning data Date 2012-11-18 Age 3 Floors 35 Floors 35 Type Residential / office Type Residential Height (m) 160 Height 160 Country UAE Country UAE Year built 2009 Year built 2009 Year refurbished n/a Under cons. No Construction type Concrete Construction Concrete Façade description Type MCM Metal composite panels consisting of aluminium with a polyethylene core. Cavity Unknown

Face Aluminium material Core PE material Insulation Unknown material Cause Reportedly cigarette discarded onto pile of waste materials possibly near to Cause Smoking_related_45-46 air conditioning equipment.

Starting floor Roof Starting floor 35 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler Sprk_Not_known_0 Probyn Miers report suggests that sprinklers were present but activation was 'delayed'.

Wind as a factor No Wind No_wind_effect

Falling material as a factor Photographs suggest downward fire spread was Falling material Increase_fire_growth partially due to molten flaming debris

Sources 1. Evans D. High-rise façade fires: a world-wide concern. Fire Safety in Towers Conference, Sept 2017. 2. White N., Delichatsios M. Fire hazards of exterior wall assemblies containing combustible components. Fire Protection Research Foundation, 2014. 3. Fire risks from external cladding panels – A perspective from the UK, Probyn Miers, 2016. www.probyn-miers.com/perspective/2016/02/fire-risks-from-external-cladding-panels- 4. Valiulisperspective-from-the-uk/ J. Building exterior wall assembly flammability, fireengineering.com, 2015. www.fireengineering.com/content/dam/fe/online-articles/documents/2015/Valiulis.pdf 5. Fire risks from external cladding panels – A perspective from the UK, Probyn Miers Ltd, 2018.

6. www.skyscrapercenter.com/building/tamweel-tower/3364

44

Name Saif Belhasa Building, Tecom ID 5 Coordinates (Longitude, latitude, elevation above sea level) (25.0937489, 55.1766079, 0) Incident description The Saif Belhasa building has 156 apartments and lower car parking level. During the fire 9 flats were damaged and falling debris damaged at least 5 cars at street level. Vertical fire spread was centred on vertical channel profiles created by balconies.

Primary database content Machine learning data Date 2012-10-06 Age Unknown Floors 13 Floors 13 Type Residential Type Residential Height (m) 57 Height 57 Country UAE Country UAE Year built Unknown Year built Unknown Year refurbished n/a Under cons. No Construction type Unknown Construction Unknown Façade description Type MCM Metal composite panels consisting of aluminium with a polyethylene core. Cavity Unknown

Face Aluminium material Core PE material Insulation Unknown material Cause Short-circuit or over heating (appears to be a guess) Cause Electric_3

Starting floor 4 Starting floor 4 Casulties Several reports state 2 injuries although one says Fatalaties 0 there were 2 deaths Injuries 2 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor No Wind No_wind_effect

Falling material as a factor Around 5 cars destroyed by falling material Falling material Caused_property_damage

Sources 1. White N., Delichatsios M. Fire hazards of exterior wall assemblies containing combustible components. Fire Protection Research Foundation, 2014. 2. Fire risks from external cladding panels – A perspective from the UK, Probyn Miers, 2016. www.probyn-miers.com/perspective/2016/02/fire-risks-from-external-cladding-panels- 3. www.gulfnews.com/news/uae/emergencies/fire-breaks-out-in-tecom-building-1.1085705

4. www.gulfbusiness.com/major-building-fires-in-dubai-this-year/

45

Name La Tour Mermoz, Roubaix ID 6 Coordinates (Longitude, latitude, elevation above sea level) (47.4079194, -1.1873486, 34) Incident description The fire started on a second floor balcony. The spread of the fire appeared to be enhanced by the profile created by the balconies and the fire also spread into the building.

Primary database content Machine learning data Date 2010-05-14 Age 7 Floors 18 Floors 18 Type Residential Type Residential Height (m) 55 Height 55 Country France Country France Year built 1966 Year built 1966 Year refurbished 2003 Under cons. No Construction type Unknown Construction Unknown Façade description Type MCM The first storey was covered with "formo-phenolic" decorative boards while Cavity Unknown the rest of the building was clad with 3 mm thick polyethylene core sandwiched between two 0.5 mm thick aluminium sheets. Face Aluminium material Core PE material Insulation Unknown material Cause Presumably of accidental Cause Unknown_0

Starting floor 2 Starting floor 2 Casulties 1 death; 6 injuries Fatalaties 1 Injuries 6 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Wind directed the fire away from the building. Wind Decrease_fire_growth

Falling material as a factor Flaming debris from the panels fell to lower Falling material Observed level balconys and the ground.

Sources 1. White N., Delichatsios M. Fire hazards of exterior wall assemblies containing combustible components. Fire Protection Research Foundation, 2014. 2. Maes P. Thoughts from a firefighter on Grenfell Tower. www.firesafeeurope.eu/fire-no-surprise- thoughts-firefighter-grenfell-tower/ 3. www.resoaplus.fr/forum/viewtopic.php?t=171

46

Name Grozny-City Towers, Groznyj ID 7 Coordinates (Longitude, latitude, elevation above sea level) (43.3159838, 45.6969238, 121) Incident description Construction had just completed in this unoccupied, high rise building. Ignition attributed to a short circuit in an air conditioner on upper floors. Fire spread to engulf the façade from ground level to the roof. Façade materials believed to be metal composite panels, but actual details not reported.

Primary database content Machine learning data Date 2013-04-03 Age 0 Floors 28 or 40 Floors 28 Type Hotel and residential Type Residential Height (m) 88 Height 88 Country Country Russia Year built 2013 Year built 2013 Year refurbished n/a Under cons. No Construction type Unknown Construction Unknown Façade description Type MCM Façade materials believed to be metal composite panels, but actual details Cavity Unknown not reported. Face Unknown material Core Unknown material Insulation Unknown material Cause Attributed to a short circuit in an air conditioner on upper floors Cause Electric_3

Starting floor Upper floor Starting floor 28 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler Sprk_Not_activated Yes, but no water supply

Wind as a factor No Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. Evans D. High-rise façade fires: a world-wide concern. Fire Safety in Towers Conference, Sept 2017. 2. White N., Delichatsios M. Fire hazards of exterior wall assemblies containing combustible components. Fire Protection Research Foundation, 2014. 3. Valiulis J. Building exterior wall assembly flammability, fireengineering.com, 2015. www.fireengineering.com/content/dam/fe/online-articles/documents/2015/Valiulis.pdf 4. www.eng.kavkaz-uzel.eu/articles/23527/

5. www.skyscrapercenter.com/building/grozny-city-residential-tower-1/15215

47

Name MGM Monte Carlo Hotel, ID 8 Coordinates (Longitude, latitude, elevation above sea level) (36.1046092, -115.1760289, 649) Incident description The fire started on the upper levels of the hotel tower due to welding. The fire spread mainly laterally with some downward spread. Exterior windows broke but the activation of 17 internal sprinklers halted any interior fire spread.

Primary database content Machine learning data Date 2008-01-25 Age 13 Floors 32 Floors 32 Type Hotel / casino Type Hotel Height (m) 110 Height 110 Country USA Country USA Year built 1995 Year built 1995 Year refurbished n/a Under cons. No Construction type Unknown Construction Unknown Façade description Type EIFS EIFS consisted of a layer of expanded polystyrene foam adhered to gypsum Cavity Unknown sheathing. Exterior the panels consisted of successive layers of fiberglass mesh and an outside coating of a weather-resistive polymer and cement Face Polymer mixture. EIFS areas had a non-complying thickness of exterior encapsulant. material Decorative non-EIFS architectural details of EPS foam encapsulated in Core n/a polyurethane resin also installed on the exterior. material Insulation EPS / XPS material Cause Welding Cause Welding_40

Starting floor Roof Starting floor 32 Casulties Thirteen guests were reportedly treated for Fatalaties 0 minor injuries and smoke inhalation. Injuries 13 Sprinklers Sprinkler Sprk_Not_known_0 Yes, in excess of NFPA 13

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Burning debris fell on evacuating occupants. Falling material Firefighting_operations

Sources 1. Evans D. High-rise façade fires: a world-wide concern. Fire Safety in Towers Conference, Sept 2017. 2. White N., Delichatsios M. Fire hazards of exterior wall assemblies containing combustible components. Fire Protection Research Foundation, 2014. 3. Duval B. Monte Carlo Hotel Casino fire, NFPA Journal, 2008

4. www.skyscrapercenter.com/building/monte-carlo-las-vegas/14153

48

Name Palace Station, Las Vegas ID 9 Coordinates (Longitude, latitude, elevation above sea level) (36.1432377, -115.1750874, 635) Incident description The fire occurred during a thunderstorm and lightning may have been the cause. The fire was confined to the outside of the building although there is a report of sprinklers activating in the building.

Primary database content Machine learning data Date 1998-07-20 Age 22 Floors 20 Floors 20 Type Hotel Type Hotel Height (m) 76 Height 76 Country USA Country USA Year built 1976 Year built 1976 Year refurbished n/a Under cons. No Construction type Unknown Construction Unknown Façade description Type EIFS Polyurethane foam and urethane coated EPS Cavity Unknown

Face Unknown material Core n/a material Insulation PU material Cause Lightning Cause Natural_occurance_52

Starting floor Unknown Starting floor Unknown Casulties 4 injuries Fatalaties 0 Injuries 4 Sprinklers Sprinkler Sprk_Not_known_0 Yes (activated)

Wind as a factor No Wind No_wind_effect

Falling material as a factor No Falling material No

Sources 1. Evans D. High-rise façade fires: a world-wide concern. Fire Safety in Towers Conference, Sept 2017. 2. White N., Delichatsios M. Fire hazards of exterior wall assemblies containing combustible components. Fire Protection Research Foundation, 2014. 3. www.lasvegassun.com/news/1998/jul/20/officials-say-lightning-likely-cause-of-palace- sta/?history 4. www.emporis.com/buildings/122229/palace-station-hotel-casino-las-vegas-nv-usa

49

Name Eldorado Hotel, Reno, Nevada ID 10 Coordinates (Longitude, latitude, elevation above sea level) (39.5294336, -119.8149948, 1373) Incident description The cause of the fire was reported to be an electrical fault on an external wall. No spread into the building was reported and no internal sprinklers activated. Large pieces of material fell off the side of the building.

Primary database content Machine learning data Date 1997-09-30 Age 1 Floors 26 Floors 26 Type Hotel Type Hotel Height (m) 82 Height 82 Country USA Country USA Year built 1973 Year built 1973 Year refurbished 1996 Under cons. No Construction type Unknown Construction Unknown Façade description Type MCM Believed to be hard coat polyurethane 'sign' over EPS to extend Cavity Unknown approximately 120 ft long and 60 ft high but no detailed description of the façade other than it being a curtain wall system. Face Polymer material Core Unknown material Insulation EPS / XPS material Cause Electrical fault Cause Electric_3

Starting floor Unknown Starting floor Unknown Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler Sprk_Not_activated Sprinkler protected but no internal sprinklers reported activated

Wind as a factor No Wind No_wind_effect

Falling material as a factor It appears the falling material did not contribute Falling material Observed to the fire development.

Sources 1. Evans D. High-rise façade fires: a world-wide concern. Fire Safety in Towers Conference, Sept 2017. 2. White N., Delichatsios M. Fire hazards of exterior wall assemblies containing combustible components. Fire Protection Research Foundation, 2014. 3. www.newspapers.com/newspage/153202396/

4. www.emporis.com/buildings/126812/eldorado-hotel-casino-reno-nv-usa

50

Name Wooshin Golden Suites, Busan ID 11 Coordinates (Longitude, latitude, elevation above sea level) (35.1565031, 129.1474966, 5) Incident description The fire appeared to have started at a janitor’s garbage collection room. A vertical 'U' shaped channel enhanced fire spread through re-radiation and chimney effect. The fire spread may also have been enhanced by a strong wind.

Primary database content Machine learning data Date 2010-10-01 Age 5 Floors 38 Floors 38 Type Mixed use (mostly apartments) Type Residential Height (m) 140 Height 140 Country Country South Korea Year built 2005 Year built 2005 Year refurbished n/a Under cons. No Construction type Steel with reinforced concrete in part Construction Steel Façade description Type MCM Aluminum composite panels with a 3 mm polyethylene core. Reports Cavity Unknown conflict about whether the thermal insulation was glass wool or EPS . Face Aluminium material Core PE material Insulation EPS / XPS material Cause Fire started on the fourth floor due to a spark from an electrical outlet. Cause Electric_3

Starting floor 4 Starting floor 4 Casulties 5 injuries (4 residents, 1 fire-fighter) Fatalaties 0 Injuries 5 Sprinklers Sprinkler Sprk_Not_activated Reported to be sprinkler protected but not to have sprinklers in the room of fire origin

Wind as a factor Wind reported to have 'fuelled' the fire Wind Increase_fire_growth

Falling material as a factor Unknown Falling material No

Sources 1. Evans D. High-rise façade fires: a world-wide concern. Fire Safety in Towers Conference, Sept 2017. 2. White N., Delichatsios M. Fire hazards of exterior wall assemblies containing combustible components. Fire Protection Research Foundation, 2014. 3. Kim Y-S, Mizuno M, Ohmiya Y. Fire examination of superhigh-rise apartment building “Wooshin Golden Suites” in Busan, Korea, Fire Science and Technology, 30(3) 81-90, 2011. 4. www.koreabridge.net/post/haeundae-highrise-fire-busan-marine-city-burns

5. www.skyscrapercenter.com/building/wooshin-golden-suites/9938

51

Name Television Cultural Centre, Mandarin Oriental ID 12 Coordinates (Longitude, latitude,Hotel, elevationBeijing above sea level) (39.9152751, 116.4642312, 44) Incident description The upper portion of the China Central Television headquarters (CCTV) facade was ignited by illegal fireworks. The fire spread to involve the majority of the facade over the entire height of building, which is believed to have included polystyrene insulation.

Primary database content Machine learning data Date 2009-02-09 Age 0 Floors 31 Floors 31 Type Hotel Type Hotel Height (m) 159 Height 159 Country China Country China Year built n/a Year built n/a Year refurbished n/a Under cons. Yes Construction type Steel Construction Steel Façade description Type EIFS Extruded insulating foam panels and polystyrene insulation have been Cavity Unknown implicated. It was also possible that incomplete construction was a factor. There was no ignition of façades insulated with mineral wool. Titanium-zinc Face Zinc alloy, EPDM rubbler water-proof sheet. material Core n/a material Insulation EPS / XPS material Cause Ignited by illegal fireworks Cause Fireworks_49

Starting floor Roof Starting floor 31 Casulties 1 fire-fighter death; 7 injuries (6 fire-fighters) Fatalaties 1 Injuries 7 Sprinklers Sprinkler Sprk_Not_activated Lack of working sprinklers (Wikipedia)

Wind as a factor Strongs winds hampered fire-fighting. Wind Firefighting_operations

Falling material as a factor Melting and dripping helped to accelerate the Falling material Increase_fire_growth fire.

Sources 1. Evans D. High-rise façade fires: a world-wide concern. Fire Safety in Towers Conference, Sept 2017. 2. White N., Delichatsios M. Fire hazards of exterior wall assemblies containing combustible components. Fire Protection Research Foundation, 2014. 3. Peng L, Ni Z, Huang X. Review on the fire safety of exterior wall claddings in high-rise buildings in China, Procedia Engineering, 62, 663–670, 2013. 4. www.nytimes.com/2009/02/11/world/asia/11beijing.html

5. www.skyscrapercenter.com/building/tvcc/3418

52

Name Center International Plaza, Nanjing City ID 13 Coordinates (Longitude, latitude, elevation above sea level) (32.0660252, 118.7771155, 24) Incident description There is very little information available on this incident. Photographs show burning material from the building.

Primary database content Machine learning data Date 2009-04-19 Age 0 Floors 50 Floors 50 Type Office Type Office Height (m) 203 Height 203 Country China Country China Year built 2009 Year built 2009 Year refurbished n/a Under cons. No Construction type Unknown Construction Unknown Façade description Type Unknown Unknown Cavity Unknown

Face Unknown material Core Unknown material Insulation Unknown material Cause Unknown Cause Unknown_0

Starting floor Unknown Starting floor Unknown Casulties Unknown Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unclear whether falling material contributed to Falling material Observed the fire development.

Sources 1. Evans D. High-rise façade fires: a world-wide concern. Fire Safety in Towers Conference, Sept 2017. 2. english.cri.cn/6909/2009/04/19/189s476237.htm

3. www.emporis.com/buildings/278785/center-international-square-nanjing-china

53

Name Baku, Binaqadi district ID 14 Coordinates (Longitude, latitude, elevation above sea level) Unknown Incident description The fire occurred in an apartment building that had been renovated as part of a wider programme started around 2012. One report states that polyurethane panels were erected nearly 30 cm away from the original concrete walls, creating a vertical tunnel. The district is also referred to as 'Binagadi' in some reports.

Primary database content Machine learning data Date 2015-05-19 Age 3 Floors 16 Floors 16 Type Residential Type Residential Height (m) Unknown Height Unknown Country Azerbaijan Country Azerbaijan Year built Unknown Year built Unknown Year refurbished 2012 Under cons. No Construction type Unknown Construction Unknown Façade description Type EIFS Reports imply the facade was made of MCM with PE core. However one Cavity Unknown video of another similar building shows an EIFS-type product being used. Face Unknown material Core n/a material Insulation Foam material Cause Unknown Cause Unknown_0

Starting floor Unknown Starting floor Unknown Casulties 15 deaths; 63 injuries Fatalaties 15 Injuries 63 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Strong winds fanned the flames. Wind Increase_fire_growth

Falling material as a factor Unknown Falling material No

Sources 1. Evans D. High-rise façade fires: a world-wide concern. Fire Safety in Towers Conference, Sept 2017. 2. www.rferl.org/a/azerbaijan-public-anger-over-deadly-fire/27027429.html

3. www.dailymail.co.uk/news/article-3088022/Tower-block-inferno-claims-16-lives-including-three- children-leaves-60-injured-blaze-rips-building-Azerbaijan-blamed-low-quality-materials.html

54

Name Baku, Zikh highway ID 15 Coordinates (Longitude, latitude, elevation above sea level) Unknown Incident description The fire occurred in an apartment building that had been renovated as part of a wider programme started around 2012. Very few further details have been found.

Primary database content Machine learning data Date 2015-04-10 Age 3 Floors 9 Floors 9 Type Residential Type Residential Height (m) Unknown Height Unknown Country Azerbaijan Country Azerbaijan Year built Unknown Year built Unknown Year refurbished 2012 Under cons. No Construction type Unknown Construction Unknown Façade description Type Other Report mentions that the façade was made of low quality inflammable Cavity Unknown material. Face Unknown material Core Unknown material Insulation Unknown material Cause Unknown Cause Unknown_0

Starting floor Unknown Starting floor Unknown Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. en.trend.az/azerbaijan/society/2382508.html

55

Name Polat Tower, ID 16 Coordinates (Longitude, latitude, elevation above sea level) (41.0569586, 28.9999466, 40) Incident description Tower contained flats, shops and businesses. The fire was fanned by strong winds and burned through the building's external insulation.

Primary database content Machine learning data Date 2012-07-17 Age 11 Floors 42 Floors 42 Type Residential Type Residential Height (m) 152 Height 152 Country Country Turkey Year built 2001 Year built 2001 Year refurbished n/a Under cons. No Construction type Unknown Construction Unknown Façade description Type Unknown Unknown Cavity Unknown

Face Unknown material Core Unknown material Insulation Unknown material Cause Faulty air conditioning unit Cause Electric_3

Starting floor Base of tower Starting floor 1 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler Sprk_Not_known_0 Reports state that building's fire extinguishing system automatically activated.

Wind as a factor Fire was fanned by strong winds. Wind Increase_fire_growth

Falling material as a factor Unknown Falling material No

Sources 1. Evans D. High-rise façade fires: a world-wide concern. Fire Safety in Towers Conference, Sept 2017. 2. www.dailymail.co.uk/news/article-2174853/Polat-Tower-Firefighters-huge-blaze-engulfed-150m- Istanbul-skyscraper.html 3. www.bbc.co.uk/news/world-europe-18867266

4. www.skyscrapercenter.com/building/polat-tower-residence/3837

56

Name Al Tayer Tower, Sharjah ID 17 Coordinates (Longitude, latitude, elevation above sea level) (25.3016451, 55.3742450, 7) Incident description The fire is thought to have started from discarded cigarette butt landing on a 1st floor balcony which contained cardboard boxes and plastics. One report suggested wind was an important factor on the incident. Fire destroyed 45 vehicles due to burning debris falling on them to ground level.

Primary database content Machine learning data Date 2012-04-28 Age 3 Floors 41 Floors 41 Type Residential Type Residential Height (m) 161 Height 161 Country UAE Country UAE Year built 2009 Year built 2009 Year refurbished n/a Under cons. No Construction type Concrete Construction Concrete Façade description Type MCM Metal composite panels consisting of aluminium with a polyethylene core Cavity Unknown

Face Aluminium material Core PE material Insulation Unknown material Cause Cigarette butt Cause Smoking_related_45-46

Starting floor 1 Starting floor 1 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Reported as an important factor Wind Increase_fire_growth

Falling material as a factor Destruction of cars Falling material Caused_property_damage

Sources 1. Evans D. High-rise façade fires: a world-wide concern. Fire Safety in Towers Conference, Sept 2017. 2. White N., Delichatsios M. Fire hazards of exterior wall assemblies containing combustible components. Fire Protection Research Foundation, 2014. 3. Valiulis J. Building exterior wall assembly flammability, fireengineering.com, 2015. www.fireengineering.com/content/dam/fe/online-articles/documents/2015/Valiulis.pdf 4. Fire risks from external cladding panels – A perspective from the UK, Probyn Miers, 2016. www.probyn-miers.com/perspective/2016/02/fire-risks-from-external-cladding-panels- 5. www.emporis.com/buildings/1286596/al-tayer-building-sharjah-united-arab-emirates

57

Name Water Club Tower, Atlantic City ID 18 Coordinates (Longitude, latitude, elevation above sea level) (39.3783325, -74.4352455, 6) Incident description The fire occurred while the building was nearing completion. The aluminium composite panels were used as a decorative finish on a structural frame set approximately 2 m distant from a concrete sheer wall that prevented major fire extension into the building. The fire started on the 3rd floor within the building. The fuel was rapidly consumed with around 10 to 15 min.

Primary database content Machine learning data Date 2007-09-23 Age 0 Floors 42 Floors 42 Type Hotel / casino Type Hotel Height (m) 131 Height 131 Country USA Country USA Year built 2007 Year built 2007 Year refurbished n/a Under cons. Yes Construction type Concrete Construction Concrete Façade description Type MCM Aluminium composite panel system composed of 3 mm aluminum sheets Cavity Unknown bonded to a 6 mm polyethylene core. The back sides of the panels were covered with 18 mm thick fire retardant polystyrene insulation. Face Aluminium material Core PE material Insulation EPS / XPS material Cause Unknown Cause Unknown_0

Starting floor 3 Starting floor 3 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler Sprk_Not_known_0 Sprinkler activations up to the 10th floor

Wind as a factor Wind prevented debris, smoke, and heat from Wind Decrease_fire_growth penetrating the building’s interior. Falling material as a factor Falling debris was a significant around the Falling material Observed building.

Sources 1. Evans D. High-rise façade fires: a world-wide concern. Fire Safety in Towers Conference, Sept 2017. 2. White N., Delichatsios M. Fire hazards of exterior wall assemblies containing combustible components. Fire Protection Research Foundation, 2014. 3. Foley J M, Modern building materials are factors in Atlantic City fires, Fire Engineering, 2010.

4. www.skyscrapercenter.com/building/borgata-hotel-casino/10266

58

Name Középszer Street, Miskolc ID 19 Coordinates (Longitude, latitude, elevation above sea level) (48.0813148, 20.7827716, 151) Incident description The fire started in a kitchen and resulted fire spread over the facade to the top of the building as well as smoke spread through stair and mechanical shafts. Invesigations found that the system was not constructed in accordance with industry requirements with lamina not adequately adhered to polystyrene sheets and there was no mineral wool insulation used as fire barriers, particularly around the windows.

Primary database content Machine learning data Date 2009-08-15 Age 2 Floors 11 Floors 11 Type Residential Type Residential Height (m) Unknown Height Unknown Country Hungary Country Hungary Year built 1968 Year built 1968 Year refurbished 2007 Under cons. No Construction type Concrete Construction Concrete Façade description Type EIFS EIFS on the exterior walls consisting of 70 mm thick polystyrene with a thin Cavity Unknown render on top. Face Render material Core n/a material Insulation EPS / XPS material Cause Insufficient evidence Cause Unknown_0

Starting floor 6 Starting floor 6 Casulties 3 deaths Fatalaties 3 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. White N., Delichatsios M. Fire hazards of exterior wall assemblies containing combustible components. Fire Protection Research Foundation, 2014. 2. Hajpal M. Analysis of a tragic fire case in panel building of Miskolc, COST TU0904 meeting, 10-11 April 2012, Malta. 3. www.irishtimes.com/news/environment/insulation-blamed-for-london-fire-widely-used-in- state-1.3121581

59

Name 393 Kennedy St, Winnipeg, Manitoba ID 20 Coordinates (Longitude, latitude, elevation above sea level) (49.8954940, -97.1499480, 232) Incident description Apartment building with open-air parking on the ground floor. Fire started in the garage involving 25 cars which then spread to the EIFS walls.

Primary database content Machine learning data Date 1990-01-10 Age 3 Floors 8 Floors 8 Type Residential Type Residential Height (m) Unknown Height Unknown Country Canada Country Canada Year built 1987 Year built 1987 Year refurbished n/a Under cons. No Construction type Unknown Construction Unknown Façade description Type EIFS EIFS typically with 75 mm thick foam but up to 140 mm. No horizontal fire Cavity Unknown barriers. Face Unknown material Core n/a material Insulation Foam material Cause Unknown Cause Unknown_0

Starting floor 1 Starting floor 1 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unsprinklered building

Wind as a factor Wind pushed flames across the car park. Wind Increase_fire_growth

Falling material as a factor Unknown Falling material No

Sources 1. White N., Delichatsios M. Fire hazards of exterior wall assemblies containing combustible components. Fire Protection Research Foundation, 2014. 2. Wade C, Clampett J. Fire performance of exterior claddings, FCRC PR 00-03, Fire Code Reform Centre Ltd, 2000.

60

Name Sonacotra building, Dijon ID 21 Coordinates (Longitude, latitude, elevation above sea level) Unknown Incident description The fire started in a garbage container at the base of the apartment building. Fire spread appeared to be influenced by the channel created by balconies as well as wind blowing flames against the wall. The fire spread smoke into the building.

Primary database content Machine learning data Date 2010-11-14 Age 23 Floors Unknown Floors 10 Type Residential Type Residential Height (m) Unknown Height Unknown Country France Country France Year built Unknown Year built 1973 Year refurbished 1987 Under cons. No Construction type Unknown Construction Unknown Façade description Type EIFS Façade believed to be EIFS system with EPS insulation and mineral wool fire Cavity Unknown barriers. Face Unknown material Core n/a material Insulation EPS / XPS material Cause External garbage container Cause Fuel_related_56

Starting floor 1 Starting floor 1 Casulties 7 deaths; 11 injuries Fatalaties 7 Injuries 11 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Wind blew flames against the wall. Wind Increase_fire_growth

Falling material as a factor Unknown Falling material No

Sources 1. White N., Delichatsios M. Fire hazards of exterior wall assemblies containing combustible components. Fire Protection Research Foundation, 2014. 2. www.bbc.co.uk/news/world-europe-11752303

3. www.lemoniteur.fr/article/incendie-du-foyer-adoma-a-dijon-les-experts-ont-rendu-leur- rapport-13577881

61

Name Treskowstraße Pankow, Berlin ID 22 Coordinates (Longitude, latitude, elevation above sea level) (52.5610340, 13.4312234, 53) Incident description Fire started in a second floor apartment with flames from a window igniting the EIFS façade. Fire and smoke spread into rooms above the starting floor.

Primary database content Machine learning data Date 2005-04-21 Age 9 Floors 7 Floors 7 Type Residential Type Residential Height (m) Unknown Height Unknown Country Germany Country Germany Year built 1996 Year built 1996 Year refurbished n/a Under cons. No Construction type Concrete poured with a lost formwork consisting ofConstruction 25 mm chipboardConcrete Façade description Type EIFS The façade consisted of 80 mm flame-retarded expanded polystyrene (EPS) Cavity None with mesh and render and mounted on 25 mm thick chipboard, which was the formwork left in place when the concrete walls had been built. In 2004 a Face Render 500 mm thick fire barrier (mineral fibre) was added to the second and fourth material levels Core n/a material Insulation EPS / XPS material Cause Unknown Cause Unknown_0

Starting floor 2 Starting floor 2 Casulties 2 deaths; 3 injuries Fatalaties 2 Injuries 3 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. White N., Delichatsios M. Fire hazards of exterior wall assemblies containing combustible components. Fire Protection Research Foundation, 2014. 2. www.irishtimes.com/news/environment/insulation-blamed-for-london-fire-widely-used-in- state-1.3121581

62

Name Switzerland ID 23 Coordinates (Longitude, latitude, elevation above sea level) Unknown Incident description A fire in a container of rubbish spread to the façade. Fire spread to upper floors through broken windows. The exact date of the fire is unknown.

Primary database content Machine learning data Date 1996-01-01 Age Unknown Floors 5 Floors 5 Type Residential Type Residential Height (m) Unknown Height Unknown Country Switzerland Country Switzerland Year built Unknown Year built Unknown Year refurbished Unknown Under cons. No Construction type Unknown Construction Unknown Façade description Type EIFS Façade made of a composite thermal insulation (about 100 mm thick) Cavity Unknown comprising polystyrene and foam plastic slabs and a reinforced covering layer Face Unknown material Core n/a material Insulation EPS / XPS material Cause Rubbish container fire on the exterior Cause Fuel_related_56

Starting floor 1 Starting floor 1 Casulties None reported Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. White N., Delichatsios M. Fire hazards of exterior wall assemblies containing combustible components. Fire Protection Research Foundation, 2014. 2. Wade C, Clampett J. Fire performance of exterior claddings, FCRC PR 00-03, Fire Code Reform Centre Ltd, 2000. 3. Meyer J. Brennender Müllcontainer gefährdet Wohngebäude, Bayerischen Versicherungskammer

63

Name Knowsley Heights, Liverpool ID 24 Coordinates (Longitude, latitude, elevation above sea level) (53.4217043, -2.8436984, 44) Incident description Fire was started in a rubbish compound and spread all the way to the highest floor, seriously damaging the outer walls and windows of all the upper floors.

Primary database content Machine learning data Date 1991-04-05 Age 1 Floors 11 Floors 11 Type Residential Type Residential Height (m) Unknown Height Unknown Country UK Country UK Year built 1960s Year built 1965 Year refurbished 1990 Under cons. No Construction type Concrete Construction Concrete Façade description Type Other Rain screen cladding installed with a 90 mm air gap and rubberised paint Cavity Ventilated coating over the external surface of the concrete wall behind. No fire barriers were provided to the air cavity. The rain screen material was a Class Face Unknown 0 (limited combustibility) rated product using BS 476 parts 6 and 7 material Core Unknown material Insulation Unknown material Cause Rubbish compound outside the building Cause Fuel_related_56

Starting floor 1 Starting floor 1 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. White N., Delichatsios M. Fire hazards of exterior wall assemblies containing combustible components. Fire Protection Research Foundation, 2014. 2. Wade C, Clampett J. Fire performance of exterior claddings, FCRC PR 00-03, Fire Code Reform Centre Ltd, 2000.

64

Name Garnock Court, Irvine ID 25 Coordinates (Longitude, latitude, elevation above sea level) (55.6110082, -4.6669923, 8) Incident description Fire began in a flat on the fifth floor and broke through a windows with the result that flames reaching the twelfth floor within 10 min. Renovation work had been carried "in the past few years" since the incident.

Primary database content Machine learning data Date 1999-06-11 Age 4 Floors 13 Floors 13 Type Residential Type Residential Height (m) Unknown Height Unknown Country UK Country UK Year built 1967 Year built 1967 Year refurbished 1995 Under cons. No Construction type Concrete Construction Concrete Façade description Type Other New uPVC window frames. Exterior wall around the window was covered Cavity Unknown with glass reinforced polyester plastic sheet. Face Polymer material Core Unknown material Insulation Unknown material Cause Room fire Cause Unknown_0

Starting floor 5 Starting floor 5 Casulties 1 death; 5 injuries Fatalaties 1 Injuries 5 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Cladding fell onto at least one fire appliance. Falling material Firefighting_operations

Sources 1. White N., Delichatsios M. Fire hazards of exterior wall assemblies containing combustible components. Fire Protection Research Foundation, 2014. 2. Wade C, Clampett J. Fire performance of exterior claddings, FCRC PR 00-03, Fire Code Reform Centre Ltd, 2000. 3. www.sundaypost.com/fp/scotlands-tragic-record-of-high-rise-fires-the-questions-raised-and- the-human-cost/ 4. www.heraldscotland.com/news/12268601.Man_dies_in_tower_block_inferno/

65

Name Ajman One, Ajman ID 26 Coordinates (Longitude, latitude, elevation above sea level) (25.3948893, 55.4307615, 0) Incident description The fire started in one tower in the complex and then spread to a neighbouring tower when winds carried burning debris toward it.

Primary database content Machine learning data Date 2016-03-28 Age 5 Floors 36 Floors 36 Type Residential Type Residential Height (m) 130 Height 130 Country UAE Country UAE Year built 2011 Year built 2011 Year refurbished n/a Under cons. No Construction type Concrete Construction Concrete Façade description Type MCM Reported to be ACM. Cavity Unknown

Face Aluminium material Core Unknown material Insulation Unknown material Cause Flaming material (possibly coal used in shisha smoking) fell from a flat and Cause Smoking_related_45-46 landed on construction waste

Starting floor 1 Starting floor 1 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Strongs winds assisted with fire spread between Wind Increase_fire_growth towers. Falling material as a factor Flaming debris seen falling to the ground. Falling material Observed

Sources 1. Evans D. High-rise façade fires: a world-wide concern. Fire Safety in Towers Conference, Sept 2017. 2. www.bbc.co.uk/news/world-middle-east-35913988

3. www.skyscrapercenter.com/complex/2461

4. www.emporis.com/buildings/332617/ajman-one-tower-6-ajman-united-arab-emirates

66

Name Sulafa Tower, Dubai ID 27 Coordinates (Longitude, latitude, elevation above sea level) (25.0893600, 55.1490086, 3) Incident description Fire started on the 36th floor.

Primary database content Machine learning data Date 2016-07-20 Age 6 Floors 76 Floors 76 Type Residential Type Residential Height (m) 288 Height 288 Country UAE Country UAE Year built 2010 Year built 2010 Year refurbished n/a Under cons. No Construction type Concrete Construction Concrete Façade description Type Unknown Unknown Cavity Unknown

Face Unknown material Core Unknown material Insulation Unknown material Cause Unknown Cause Unknown_0

Starting floor 35 Starting floor 35 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Fire spread was aided by flaming debris falling Falling material Increase_fire_growth onto lower level balconies.

Sources 1. www.dailymail.co.uk/news/article-3699272/Fire-breaks-luxury-75-storey-tower-Dubai.html

2. http://www.skyscrapercenter.com/building/sulafa-tower/606

67

Name Sulafa Tower, Dubai ID 28 Coordinates (Longitude, latitude, elevation above sea level) (25.0893600, 55.1490086, 3) Incident description Residents were forced to evacuate from the fire that started on the 36th floor early in the morning.

Primary database content Machine learning data Date 2012-06-08 Age 2 Floors 76 Floors 76 Type Residential Type Residential Height (m) 288 Height 288 Country UAE Country UAE Year built 2010 Year built 2010 Year refurbished n/a Under cons. No Construction type Concrete Construction Concrete Façade description Type Unknown Unknown Cavity Unknown

Face Unknown material Core Unknown material Insulation Unknown material Cause Unknown Cause Unknown_0

Starting floor 36 Starting floor 36 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. gulfbusiness.com/major-building-fires-in-dubai-this-year/

2. http://www.skyscrapercenter.com/building/sulafa-tower/606

68

Name Hiroshima Motomachi ID 29 Coordinates (Longitude, latitude, elevation above sea level) (34.4032152, 132.4554643, 9) Incident description The incident involved a fast spreading fire spread owing to PMMA fences used in the balconies of the building. The fire development was due to a combination of the PMMA plus additional combustibles located on the balconies. The fire caused total or partial damage to 27 apartments.

Primary database content Machine learning data Date 1996-10-28 Age 24 Floors 20 Floors 20 Type Residential Type Residential Height (m) Unknown Height Unknown Country Japan Country Japan Year built 1972 Year built 1972 Year refurbished Unknown Under cons. No Construction type Steel reinforced concrete Construction Concrete Façade description Type Other 1 m tall, 8-20 mm thick PMMA blindfold doors along balconies Cavity Unknown

Face Unknown material Core Polymer material Insulation Unknown material Cause Unknown Cause Unknown_0

Starting floor 9 Starting floor 9 Casulties Unknown Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Wind described as weak at 1 m/s Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. White N., Delichatsios M. Fire hazards of exterior wall assemblies containing combustible components. Fire Protection Research Foundation, 2014. 2. Hokugo A, Hasemi Y, Hayashi Y, Yoshida M. Mechanism for the upward fire spread through balconies based on an investigation and experiments for a multi-story fire in high-rise apartment

69

Name Manchester, New Hampshire ID 30 Coordinates (Longitude, latitude, elevation above sea level) Unknown Incident description The incident involved a flash fire on the exterior of the building due to exposure to radiant heat. The fire was extinguished in a few minutes.

Primary database content Machine learning data Date 2005-06-07 Age Unknown Floors 7 Floors 7 Type Office Type Office Height (m) Unknown Height Unknown Country USA Country USA Year built Unknown Year built Unknown Year refurbished Unknown Under cons. No Construction type Unknown Construction Unknown Façade description Type EIFS EIFS, no further information has been found. Cavity Unknown

Face Unknown material Core Unknown material Insulation Unknown material Cause External radiation Cause Fuel_related_56

Starting floor Unknown Starting floor Unknown Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. Wade C, Clampett J. Fire performance of exterior claddings, FCRC PR 00-03, Fire Code Reform Centre Ltd, 2000.

70

Name Lakeside Plaza, Virginia ID 31 Coordinates (Longitude, latitude, elevation above sea level) Unknown Incident description The fire started in a rubbish chute. Extensive vertical spread did not occur so that damage was limited to floors above and areas that had been exposed to flame heating.

Primary database content Machine learning data Date 2005-06-05 Age 22 Floors 12 Floors 12 Type Residential Type Residential Height (m) 45 Height 45 Country USA Country USA Year built 1983 Year built 1983 Year refurbished n/a Under cons. No Construction type Concrete Construction Unknown Façade description Type EIFS EIFS system with 25 mm thick XPS backing Cavity Unknown

Face Unknown material Core n/a material Insulation EPS / XPS material Cause Fire started in a trash chute on the fifth or sixth floor Cause Fuel_related_56

Starting floor 5th or 6th floor Starting floor 5 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. Wade C, Clampett J. Fire performance of exterior claddings, FCRC PR 00-03, Fire Code Reform Centre Ltd, 2000. 2. www.lakesideplaza.com/about-lp/

3. www.emporis.com/buildings/323877/lakeside-plaza-baileys-crossroads-va-usa

71

Name Country Comfort Motel, Albury, NSW ID 32 Coordinates (Longitude, latitude, elevation above sea level) (-36.0802414, 146.9125463, 160) Incident description The fire started in a third-floor apartment and when flames broke through a window there was rapid spread to the top of the building. The fire entered the building at the fourth level. The building is currently known as The Marque.

Primary database content Machine learning data Date 2005-06-20 Age 25 Floors 8 Floors 8 Type Motel Type Hotel Height (m) 42 Height 42 Country Country Australia Year built 1980 Year built 1980 Year refurbished n/a Under cons. No Construction type Concrete/masonry construction Construction Concrete Façade description Type EIFS Fibreglass panels beneath the windows. Cavity Unknown

Face Unknown material Core n/a material Insulation Mineral wool material Cause Unknown Cause Unknown_0

Starting floor 3 Starting floor 3 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. Wade C, Clampett J. Fire performance of exterior claddings, FCRC PR 00-03, Fire Code Reform Centre Ltd, 2000. 2. www.emporis.com/buildings/186364/the-marque-sydney-australia

72

Name Butler House, Grays, Essex ID 33 Coordinates (Longitude, latitude, elevation above sea level) (51.4724929, 0.3233527, 3) Incident description Fire caused uPVC window frames to melt and drip, which then led to damage to the cladding. The building was at least 40 years old in 2015 suggesting it was built on or before 1975.

Primary database content Machine learning data Date 1997-01-01 Age 22 Floors 14 Floors 14 Type Residential Type Residential Height (m) Unknown Height Unknown Country UK Country UK Year built 1975 Year built 1975 Year refurbished n/a Under cons. No Construction type Unknown Construction Unknown Façade description Type MCM Unknown Cavity Unknown

Face Unknown material Core Unknown material Insulation Unknown material Cause Unknown Cause Unknown_0

Starting floor 14 Starting floor 14 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system No

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. Potential risk of fire spread in buildings via external cladding systems; Environment, Transport and Regional Affairs Committee, House of Commons session, 1999. 2. www.braintreeandwithamtimes.co.uk/news/south_essex_news/11880093.Could_these_iconic_G rays_tower_blocks_be_coming_down__Residents_tell_of_/ 3. Essex County Fire and Rescue Service sprinkler funding report, Appendix A

73

Name Lacrosse Tower, Melbourne ID 34 Coordinates (Longitude, latitude, elevation above sea level) (-37.8147958, 144.9477586, 13) Incident description A discarded cigarette started a fire on an eigth-floor balcony which led to rapid flame spread over the building cladding. The entire population of around 400 people were evacuated from the building. The internal sprinkler system limited fire spread even though it was operating well beyond its designed capability.

Primary database content Machine learning data Date 2014-11-25 Age 2 Floors 21 Floors 21 Type Residential Type Residential Height (m) 72 Height 72 Country Australia Country Australia Year built 2012 Year built 2012 Year refurbished n/a Under cons. No Construction type Masonry, concrete and dry Wall Construction Concrete Façade description Type MCM Polyethylene core aluminium cladding Cavity Unknown

Face Aluminium material Core PE material Insulation Unknown material Cause Cigarette Cause Smoking_related_45-46

Starting floor 8 Starting floor 8 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler Sprk_Contained_2 Yes - worked

Wind as a factor Wind likely assisted in drawing flames away Wind Decrease_fire_growth from internal building Falling material as a factor Embers and flaming debris ignted materials on Falling material Increase_fire_growth Level 6 balcony

Sources 1. Melbourne Fire Brigade, Lacrosse Doclands, Post incident analysis, Report 1403134A, undated

2. www.skyscrapercenter.com/building/lacrosse-apartments/23697

74

Name Al Baker Tower 4, Sharjah ID 35 Coordinates (Longitude, latitude, elevation above sea level) (25.3083484, 55.3724543, 11) Incident description The building's exterior was made of flammable materials and the weather and wind speed were major factors that caused the fire to spread quickly to other floors in the building. Caused by a lit cigarette that was thrown off the balcony from an upper floor and landed on the balcony on the first floor. Sharjah Civil Defence found the quality of construction materials to be the main factor in the fire spreading rapidly across the rear façade

Primary database content Machine learning data Date 2012-01-18 Age 2 Floors 29 Floors 29 Type Residential Type Residential Height (m) Unknown Height Unknown Country UAE Country UAE Year built 2010 Year built 2010 Year refurbished n/a Under cons. No Construction type Concrete Construction Concrete Façade description Type Unknown Unknown Cavity Unknown

Face Unknown material Core Unknown material Insulation Unknown material Cause Lit cigarette that was thrown off the balcony Cause Smoking_related_45-46

Starting floor 1 Starting floor 1 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unsprinklered building

Wind as a factor Wind said to be a major factor that caused rapid Wind Increase_fire_growth fire spread Falling material as a factor Unknown Falling material No

Sources 1. gulfnews.com/news/uae/emergencies/tossed-lighted-cigarette-caused-fire-in-al-baker-tower- 1.1012778

75

Name Neo Soho, ID 36 Coordinates (Longitude, latitude, elevation above sea level) (-6.1746154, 106.7899857, 4) Incident description Fire in a combined retail, parking and residential building still under construction. The fire started on the ninth-level parking floor.

Primary database content Machine learning data Date 2017-11-09 Age 0 Floors 38 Floors 38 Type Residential / retail Type Residential Height (m) Unknown Height Unknown Country Country Indonesia Year built n/a Year built n/a Year refurbished n/a Under cons. Yes Construction type Unknown Construction Unknown Façade description Type Unknown Unknown Cavity Unknown

Face Unknown material Core Unknown material Insulation Unknown material Cause Unknown Cause Unknown_0

Starting floor 9 Starting floor 9 Casulties 1 injury Fatalaties 0 Injuries 1 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Large falling debris endangered fire-fighters Falling material Firefighting_operations

Sources 1. Petrus P P, A perspective on high rise building fires involving the façade, Asia Pacific Fire, 2017

2. en.tempo.co/read/news/2016/11/10/057819316/Police-Question-5-Witnesses-over-Neo-Soho- Fire 3. coconuts.co/jakarta/news/massive-fire-crawls-multiple-floors-uninhabited-neo-soho-apartment- building-one-worker/

76

Name Ramat Gan, Tel Aviv ID 37 Coordinates (Longitude, latitude, elevation above sea level) (32.0868969, 34.8073711, 30) Incident description The fire started in an apartment and spread quickly along the building exterior. Ocupants were evacuated using ladders and cranes. Ten floors were affected by the fire.

Primary database content Machine learning data Date 2016-06-13 Age Unknown Floors 13 Floors 13 Type Residential Type Residential Height (m) Unknown Height Unknown Country Israel Country Israel Year built Unknown Year built Unknown Year refurbished Unknown Under cons. No Construction type Unknown Construction Unknown Façade description Type Other Polyethylene (ploythene [sic.]) composite cladding Cavity Unknown

Face Unknown material Core PE material Insulation Unknown material Cause Unknown Cause Unknown_0

Starting floor 2 Starting floor 2 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Burning debris destroyed several cars Falling material Caused_property_damage

Sources 1. www.timesofisrael.com/hundreds-israeli-buildings-as-vulnerable-as--grenfell-tower/

2. www.timesofisrael.com/firefighters-battle-blaze-in-ramat-gan-building/

77

Name Grenfell Tower, London ID 38 Coordinates (Longitude, latitude, elevation above sea level) (51.5140585, -0.2158184, 8) Incident description The fire started in an electrical appliance in a fourth floor apartment before spreading to the outside of the building. The event caused extensive damage and casualties.

Primary database content Machine learning data Date 2017-06-14 Age 2 Floors 24 Floors 24 Type Residential Type Residential Height (m) 67 Height 67 Country UK Country UK Year built 1974 Year built 1974 Year refurbished 2015 Under cons. No Construction type Concrete Construction Concrete Façade description Type MCM Aluminium sheets with a polyethylene core and PIR thermal insulation. Cavity Ventilated

Face Aluminium material Core PE material Insulation PIR material Cause Electrical Cause Electric_3

Starting floor 4 Starting floor 4 Casulties 71 fatalities; 74 injuries Fatalaties 71 Injuries 74 Sprinklers Sprinkler No_sprk_system No

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Fire-fighters had to be protected from falling Falling material Firefighting_operations debris

Sources 1. www.maria-online.com/travel/article.php?lg=en&q=Grenfell_Tower_fire

2. Note: The above link provides a summary of material and an extensive reference list.

78

Name Belleroche, Villefranche-sur-Saone ID 39 Coordinates (Longitude, latitude, elevation above sea level) (45.9816661, 4.7038746, 217) Incident description A fire broke out in a stock of bulky goods, which had been stored at the foot of the building. The fire moved into building due to window failure and damaged 8 apartments.

Primary database content Machine learning data Date 2015-08-17 Age Unknown Floors 7 Floors 7 Type Residential Type Residential Height (m) Unknown Height Unknown Country France Country France Year built Unknown Year built Unknown Year refurbished Unknown Under cons. No Construction type Unknown Construction Unknown Façade description Type Unknown Unknown Cavity Unknown

Face Unknown material Core Unknown material Insulation Unknown material Cause Unknown Cause Unknown_0

Starting floor 1 Starting floor 1 Casulties 20 injuries Fatalaties 0 Injuries 20 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. www.leprogres.fr/rhone/2015/08/17/le-feu-se-propage-a-la-facade-panique-dans-un-immeuble- de-villefranche 2. firesafeeurope.eu/fire-no-surprise-thoughts-firefighter-grenfell-tower/

79

Name Blaise-Pascal, Montmarin district, Vesoul ID 40 Coordinates (Longitude, latitude, elevation above sea level) Unknown Incident description The incident was a garbage fire that broke out in front of the building. The fire destroyed a section of the newly renovated façade but was rapidly controlled by the fire service. Building was at least 15 years old at the time of the incident and a general history of the area suggests the building was constructed between 1967 and 1973.

Primary database content Machine learning data Date 2015-09-29 Age 3 Floors Unknown Floors 17 Type Residential Type Residential Height (m) Unknown Height Unknown Country France Country France Year built Likley between 1967 and 1973 Year built 1970 Year refurbished 2012 Under cons. No Construction type Unknown Construction Unknown Façade description Type Unknown Unknown Cavity Unknown

Face Unknown material Core Unknown material Insulation Unknown material Cause Unknown Cause Unknown_0

Starting floor 1 Starting floor 1 Casulties 2 hospitalisations Fatalaties 0 Injuries 2 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. www.estrepublicain.fr/edition-de-vesoul-haute-saone/2015/09/30/vesoul-le-feu-de-poubelles- degenere-dans-le-quartier-du-montmarin 2. firesafeeurope.eu/fire-no-surprise-thoughts-firefighter-grenfell-tower/

3. www.estrepublicain.fr/haute-saone/2011/07/16/appartement-en-feu

80

Name Lomond House, Charles Street, Glasgow ID 41 Coordinates (Longitude, latitude, elevation above sea level) (55.8714408, -4.2248795, 62) Incident description Fire spread up through eight storeys of the building. Eye witness reports suggest that cladding contributed to the incident.

Primary database content Machine learning data Date 2015-05-24 Age 54 Floors 20 Floors 20 Type Residential Type Residential Height (m) 58 Height 58 Country UK Country UK Year built 1961 Year built 1961 Year refurbished 2011 or 2013 Under cons. No Construction type Unknown Construction Unknown Façade description Type Unknown Unknown Cavity Unknown

Face Unknown material Core Unknown material Insulation Unknown material Cause Unknown Cause Unknown_0

Starting floor About half way, ~12 from video Starting floor 12 Casulties 3 injuries; 5 hospitalisations Fatalaties 0 Injuries 8 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Dripping down to flats below Falling material Increase_fire_growth

Sources 1. stv.tv/news/features/1391320-london-fire-concerns-over-cladding-stretch-back-years/

2. www.sundaypost.com/fp/scotlands-tragic-record-of-high-rise-fires-the-questions-raised-and- the-human-cost/ 3. fr.wikipedia.org/wiki/Montmarin_(Vesoul)

4. www.emporis.com/buildings/146712/160-charles-street-glasgow-united-kingdom

81

Name Red Road flats, Glasgow ID 42 Coordinates (Longitude, latitude, elevation above sea level) (55.8798320, -4.2104180, 68) Incident description Fire was discovered in the external cladding. Building was demolished in 2015.

Primary database content Machine learning data Date 2014-09-24 Age 48 Floors 31 Floors 31 Type Residential Type Residential Height (m) 89 Height 89 Country UK Country UK Year built 1966 Year built 1966 Year refurbished n/a Under cons. No Construction type Steel Construction Steel Façade description Type Unknown Unknown Cavity Unknown

Face Unknown material Core Unknown material Insulation Unknown material Cause Unknown Cause Unknown_0

Starting floor Unknown Starting floor Unknown Casulties 1 injury Fatalaties 0 Injuries 1 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. stv.tv/news/features/1391320-london-fire-concerns-over-cladding-stretch-back-years/

2. www.emporis.com/buildings/110796/10-red-road-court-glasgow-united-kingdom

82

Name BRE Case 3 ID 43 Coordinates (Longitude, latitude, elevation above sea level) Unknown Incident description Damage façade localised to the immediate vicinity of some of the windows but beyond this the damage appeared to have been limited to surface charring and sooting

Primary database content Machine learning data Date 2008-02-01 Age 1 Floors 22 Floors 22 Type Residential Type Residential Height (m) Unknown Height Unknown Country UK Country UK Year built ~1965 Year built 1965 Year refurbished 2007 Under cons. No Construction type Unknown Construction Unknown Façade description Type EIFS Insulated render - mineral wool insulation conforming to BS 8414-1 Cavity Unknown

Face Render material Core n/a material Insulation Mineral wool material Cause Unknown Cause Unknown_0

Starting floor 11 Starting floor 11 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. Holland C, Crowder D, Shipp M. External fire spread- Part 1 background research, BRE Global, Garston, UK, 2016.

83

Name BRE Case 2 ID 44 Coordinates (Longitude, latitude, elevation above sea level) Unknown Incident description Several factors contributed to this fire including combustible panels on the underside of balconies, outside cladding panels and plastic drain pipes. No fire stopping, seal or steeve was found where the pipe passed through the concrete floor slab.

Primary database content Machine learning data Date 2010-07-01 Age 45 Floors 16 Floors 16 Type Residential Type Residential Height (m) Unknown Height Unknown Country UK Country UK Year built ~1965 Year built 1965 Year refurbished Unknown Under cons. No Construction type Concrete Construction Concrete Façade description Type EIFS Cladding panels had fibrous combustible insulation material behind mineral Cavity Unknown fibre board face supported on timber batons Face Mineral fibre board material Core n/a material Insulation Combustible material Cause Unknown Cause Unknown_0

Starting floor 12 Starting floor 12 Casulties Unknown Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. Holland C, Crowder D, Shipp M. External fire spread- Part 1 background research, BRE Global, Garston, UK, 2016.

84

Name Al Nasser Tower, Sharjah ID 45 Coordinates (Longitude, latitude, elevation above sea level) (25.3297861, 55.3941289, 9) Incident description Fire occurred in a combined residential and parking building. The right side of the building was severely affected and around 40 apartments were damaged. Reports suggest that the building had been built in 2015 not using "anti-fire material for cladding but ... was built before this specification came into force" although the UAE Civil Defence Fire Code had been revised in 2012.

Primary database content Machine learning data Date 2015-10-01 Age 0 Floors 36 Floors 36 Type Residential Type Residential Height (m) Unknown Height Unknown Country UAE Country UAE Year built 2015 Year built 2015 Year refurbished n/a Under cons. No Construction type Concrete Construction Concrete Façade description Type Unknown Unknown Cavity Unknown

Face Unknown material Core Unknown material Insulation Unknown material Cause Unknown Cause Unknown_0

Starting floor 3 Starting floor 3 Casulties 10 injuries Fatalaties 0 Injuries 10 Sprinklers Sprinkler No_sprk_system Unsprinklered building

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Small pieces of burning debris drifted to the Falling material Caused_property_damage ground damaging seven cars.

Sources 1. gulfnews.com/news/uae/emergencies/video-massive-fire-erupts-in-sharjah-high-rise-tower- 1.1593293 2. www.arabianbusiness.com/sharjah-tower-residents-return-home-after-fire-614375.html

3. www.khaleejtimes.com/sharjah-apartments-face-inspecion-after-massive-fire

4. www.skyscrapercenter.com/building/nasser-tower/22209

85

Name Al Nasr Tower, Doha ID 46 Coordinates (Longitude, latitude, elevation above sea level) (25.3131655, 51.5195471, 10) Incident description Second of two fires occurred while the building was under construction. It is not clear whether the first fire involved the cladding. The cause was attributed to the use of banned construction materials, negligence and inexperience of workers. One report suggests that the second tower was ignited by falling burning panels from the first tower.

Primary database content Machine learning data Date 2006-05-28 Age 2006 Floors 24 Floors 24 Type Residential Type Residential Height (m) 100 Height 100 Country Qatar Country Qatar Year built n/a Year built n/a Year refurbished n/a Under cons. Yes Construction type Unknown Construction Unknown Façade description Type MCM Polyethylene core aluminium curtain wall Cavity Unknown

Face Aluminium material Core PE material Insulation Unknown material Cause Unknown Cause Unknown_0

Starting floor 1 Starting floor 1 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Strong winds hampered fire-fighting Wind Firefighting_operations

Falling material as a factor Ignition of the second tower Falling material Increase_fire_growth

Sources 1. www.scribd.com/document/66790105/AluminiumCompositePanels-WorldArchitecture-Dec2009

2. gulfnews.com/news/gulf/qatar/doha-tower-blaze-due-to-violation-of-safety-rules-1.238725

3. www.thenational.ae/business/property/most-fire-resistant-panels-still-being-ignored-for-uae- towers-despite-spate-of-blazes-1.207550 4. www.emporis.com/buildings/1152987/al-nasr-towers-north-doha-qatar

86

Name Damantou Xintiote Shopping mall, Qiqihar ID 47 Coordinates (Longitude, latitude, elevation above sea level) Unknown Incident description The fire involved billboards on the shopping mall level façade of the building. The fire was extinguished by the local .

Primary database content Machine learning data Date 2008-11-19 Age 0 Floors 32 Floors 32 Type Shopping / residential Type Commecial Height (m) Unknown Height Unknown Country China Country China Year built 2008 Year built 2008 Year refurbished n/a Under cons. No Construction type Unknown Construction Unknown Façade description Type Unknown Unknown Cavity Unknown

Face Unknown material Core Unknown material Insulation Unknown material Cause Unknown Cause Unknown_0

Starting floor 1 Starting floor 1 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. Valiulis J. Building exterior wall assembly flammability, fireengineering.com, 2015. www.fireengineering.com/content/dam/fe/online-articles/documents/2015/Valiulis.pdf 2. cn.epochtimes.com/b5/8/11/21/n2336786.htm

87

Name Jingwei 360, Harbin ID 48 Coordinates (Longitude, latitude, elevation above sea level) (45.7666436, 126.6235949, 124) Incident description The fire occurred while the building was under construction. Workers ignited construction materials while welding.

Primary database content Machine learning data Date 2008-10-10 Age 0 Floors 29 Floors 29 Type Residential Type Residential Height (m) 99 Height 99 Country China Country China Year built n/a Year built n/a Year refurbished n/a Under cons. Yes Construction type Unknown Construction Unknown Façade description Type Unknown Unknown Cavity Unknown

Face Unknown material Core Unknown material Insulation Unknown material Cause Welding Cause Welding_40

Starting floor 3 Starting floor 3 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. Valiulis J. Building exterior wall assembly flammability, fireengineering.com, 2015. www.fireengineering.com/content/dam/fe/online-articles/documents/2015/Valiulis.pdf 2. english.sina.com/china/p/2008/1009/190919.html

3. www.emporis.com/buildings/349172/jingwei-360-harbin-china

4. www.scmp.com/article/655621/fire-engulfs-unfinished-high-rise

88

Name Royal Wanxin, Tower B ID 49 Coordinates (Longitude, latitude, elevation above sea level) (41.7546500, 123.4376568, 45) Incident description The fire started due to fireworks

Primary database content Machine learning data Date 2011-02-03 Age 2 Floors 38 Floors 38 Type Residential Type Residential Height (m) 152 Height 152 Country China Country China Year built 2009 Year built 2009 Year refurbished n/a Under cons. No Construction type Unknown Construction Unknown Façade description Type MCM Façade clad in ACM with XPS insulation and a 190 - 600 mm wide cavity. The Cavity Ventilated EPS had Class B2 combustibility to Chinese standard GB 8624-1997. Face Aluminium material Core PE material Insulation EPS / XPS material Cause Fireworks Cause Fireworks_49

Starting floor 11 Starting floor 11 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. Peng L, Ni Z, Huang X. Review on the fire safety of exterior wall claddings in high-rise buildings in China, Procedia Engineering, 62, 663–670, 2013. 2. www.skyscrapercenter.com/building/shenyang-royal-wan-xin-international-mansion-tower- b/9000

89

Name Shepherds Court, London ID 50 Coordinates (Longitude, latitude, elevation above sea level) (51.5033820, -0.2194147, 6) Incident description The building was retrofitted with a new cladding around 2006. Fire started in a piece of faulty whiteware equipment in a kitchen. Flames extended through a window and spread up the outside of the building.

Primary database content Machine learning data Date 2016-08-19 Age 10 Floors 18 Floors 18 Type Residential Type Residential Height (m) 58 Height 58 Country UK Country UK Year built 1971 Year built 1971 Year refurbished 2006 Under cons. No Construction type Unknown Construction Unknown Façade description Type EIFS Panels comprised a 17-23 mm thick plywood board, covered by blue Cavity Unknown polystyrene foam, 1 mm steel sheet and decorative white paint Face Steel material Core Unknown material Insulation EPS / XPS material Cause Faulty tumble dryer Cause Electric_3

Starting floor 7 Starting floor 7 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. www.insidehousing.co.uk/insight/insight/a-stark-warning-the-shepherds-bush-tower-block-fire- 50566 2. www.emporis.com/buildings/138931/shepherds-court-london-united-kingdom

90

Name Tamworth Estate, Trafford ID 51 Coordinates (Longitude, latitude, elevation above sea level) (53.4618886, -2.2548072, 34) Incident description Raven, Osprey, Falcon and Eagle Courts. The buildings were retrofited with a new cladding in the mid-1990s. Since then four fires have occurred within two of the blocks. The fires were contained locally with smoke damage to panels.

Primary database content Machine learning data Date Unknown Age Unknown Floors 15 Floors 15 Type Residential Type Residential Height (m) 45 Height 45 Country UK Country UK Year built 1970 Year built 1970 Year refurbished 1995 Under cons. No Construction type Unknown Construction Unknown Façade description Type Other Class 0 panels with horizontal and vertical cavity fire barriers at each floor Cavity Ventilated level and between each flat. Cladding terminated at first floor. Face Unknown material Core Unknown material Insulation Unknown material Cause Unknown Cause Unknown_0

Starting floor Unknown Starting floor Unknown Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. Potential risk of fire spread in buildings via external cladding systems; Environment, Transport and Regional Affairs Committee, House of Commons session, 1999. 2. www.emporis.com/complex/127894/tamworth-estate-manchester-united-kingdom

91

Name Millennium Business Center, Bucharest ID 52 Coordinates (Longitude, latitude, elevation above sea level) (44.4373306, 26.1102233, 80) Incident description The fire was thought to have started after the building was struck by lightning.

Primary database content Machine learning data Date 2009-06-26 Age 3 Floors 19 Floors 19 Type Office Type Office Height (m) 72 Height 72 Country Romania Country Romania Year built 2006 Year built 2006 Year refurbished n/a Under cons. No Construction type Unknown Construction Unknown Façade description Type MCM ACM with 3 mm thick polyethylene core bonded between 0.5 mm Cavity Unknown aluminium facing and backing sheets. Face Aluminium material Core PE material Insulation Unknown material Cause Lightning Cause Natural_occurance_52

Starting floor Unknown Starting floor Unknown Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unsprinklered building

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. www.scribd.com/document/66790105/AluminiumCompositePanels-WorldArchitecture-Dec2009

2. www.skyscrapercenter.com/building/millenium-business-center/11303

92

Name Taksim Hospital, Istanbul ID 53 Coordinates (Longitude, latitude, elevation above sea level) (41.0888769, 28.9089101, 76) Incident description The fire reportedly started on the outside of a building under construction. Smoke entered the building and patients were evacuated from neighbouring buildings.

Primary database content Machine learning data Date 2018-04-05 Age 0 Floors Unknown Floors 11 Type Hospital Type Medical Height (m) Unknown Height Unknown Country Turkey Country Turkey Year built 2013 Year built 2013 Year refurbished n/a Under cons. Yes Construction type Unknown Construction Unknown Façade description Type Unknown Unknown Cavity Unknown

Face Unknown material Core Unknown material Insulation Unknown material Cause Unknown Cause Unknown_0

Starting floor Roof Starting floor Unknown Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Burning panels seen falling to the ground Falling material Observed

Sources 1. edition.cnn.com/2018/04/05/europe/istanbul-hospital-fire-intl/index.html

2. www.valuewalk.com/2018/04/taksim-hospital-istanbul-fire/

93

Name Târgu Mureș, Bucharest ID 54 Coordinates (Longitude, latitude, elevation above sea level) (46.5385862, 24.5514392, 309) Incident description The fire began in an outside ground floor warehouse. Flames spread to the top floors in less tan 10 min. Failure to follow manufacturer's installation specifications and inadequate bonding of polystyrene plates contributed to the incident.

Primary database content Machine learning data Date 2012-07-07 Age Unknown Floors Unknown Floors 10 Type Residential Type Residential Height (m) Unknown Height Unknown Country Romania Country Romania Year built Unknown Year built Unknown Year refurbished Unknown Under cons. No Construction type Unknown Construction Unknown Façade description Type EIFS External thermal insulation composite systems (ETICS) cladding system with Cavity Unknown polystyrene insulation Face Unknown material Core n/a material Insulation EPS / XPS material Cause Unknown Cause Unknown_0

Starting floor 1 Starting floor 1 Casulties Unknown Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Burning droplets observed Falling material Observed

Sources 1. Lalu O, Behaviour of thermal rehabilitated façades in case of fire, BRE Global presentation at the Structures in Fire Forum, 2017 2. Lalu O, Anghelb I, Șerbanb M, Mocioib I-A, Branisteanuc B. Experimental researches on determining the fire action response of improved exterior cladding systems provided with 3. www.romania-insider.com/bucharest-apartment-fire-no-injuries-but-residents-demand-refunds- deem-building-unsafe-and-refuse-to-return/

94

Name Daebong Green Apartments, Uijeongbu ID 55 Coordinates (Longitude, latitude, elevation above sea level) (37.7380980, 127.0336819, 61) Incident description The fire spread through the apartment complex and to two other adjacent residential buildings. Reports suggest that the flammable material was a 'styrofoam' product.

Primary database content Machine learning data Date 2015-01-10 Age 2 Floors 10 Floors 10 Type Residential Type Residential Height (m) Unknown Height Unknown Country South Korea Country South Korea Year built 2013 Year built 2013 Year refurbished n/a Under cons. No Construction type Unknown Construction Unknown Façade description Type Other Flammable 'styrofoam' material Cavity Unknown

Face Unknown material Core Unknown material Insulation Unknown material Cause Motorcycle spark Cause Electric_3

Starting floor 1 Starting floor 1 Casulties 4 fatalities; 100 injuries Fatalaties 4 Injuries 100 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Down-draught from helicopters blamed by some Wind No_wind_effect witnesses for spreading the fire Falling material as a factor Unknown Falling material No

Sources 1. www.koreabang.com/2015/stories/fire-in-seoul-suburb-kills-4-residents-blame-firefighters.html

2. www.koreaherald.com/view.php?ud=20150111000330

3. www.koreatimes.co.kr/www/opinion/2018/04/202_171650.html

95

Name Novaya Vysota, Krasnoyarsk ID 56 Coordinates (Longitude, latitude, elevation above sea level) (56.0394794, 92.8936420, 199) Incident description The building is also referred to as the 'New Pinnacle'. The fire spread on two sides of the building to the entire height and 14 apartments were completely burnt out.

Primary database content Machine learning data Date 2014-09-21 Age 4 Floors 25 Floors 25 Type Residential Type Residential Height (m) 86 Height 86 Country Russia Country Russia Year built 2010 Year built 2010 Year refurbished n/a Under cons. No Construction type Unknown Construction Unknown Façade description Type Other Ventilated vinyl siding and plastic-furnished balconies Cavity Ventilated

Face Polymer material Core Unknown material Insulation Unknown material Cause Welding Cause Welding_40

Starting floor Lower levels Starting floor 1 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. siberiantimes.com/other/others/news/shocking-blaze-rips-through-25-floor-apartment-block-in- krasnoyarsk/ 2. www.pravdareport.com/news/hotspots/22-09-2014/128574-fire_high_rise_building- 0/#.VCAgQxZjInE 3. www.skyscrapercenter.com/building/novaya-vysota-1/17731

4. www.emporis.com/buildings/323369/novaya-vysota-i-krasnoyarsk-russia

96

Name Fitness centre, Jecheon ID 57 Coordinates (Longitude, latitude, elevation above sea level) Unknown Incident description The fire started in the first-flor car park level. Reports suggest the cladding was a factor in the development of the incident although it is not clear what the cladding material was.

Primary database content Machine learning data Date 2017-12-22 Age 5 Floors 8 Floors 8 Type Commercial Type Commecial Height (m) Unknown Height Unknown Country South Korea Country South Korea Year built 2012 Year built 2012 Year refurbished n/a Under cons. No Construction type Unknown Construction Unknown Façade description Type EIFS Cladding materials made of a cement and foam sandwich Cavity Unknown

Face Render material Core n/a material Insulation Foam material Cause Electrical spark Cause Electric_3

Starting floor 1 Starting floor 1 Casulties 29 fatalities; 29 injuries Fatalaties 29 Injuries 29 Sprinklers Sprinkler Sprk_Not_activated One report suggests sprinklers failed

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. www.telegraph.co.uk/news/2017/12/22/deadly-south-korea-blaze-has-echoes-grenfell/

2. www.koreaherald.com/view.php?ud=20171225000179

3. english.yonhapnews.co.kr/news/2017/12/23/0200000000AEN20171223000800315.html

97

Name Ismail St, Chisinau ID 58 Coordinates (Longitude, latitude, elevation above sea level) Unknown Incident description The fire was as the result of cooking fat starting a fire in a smoke ventilation system. The fire spread to building exterior and flames travelled 70 m vertically.

Primary database content Machine learning data Date 2016-01-23 Age 4 Floors 6 Floors 6 Type Retail Type Commecial Height (m) Unknown Height Unknown Country Moldova Country Moldova Year built 2012 Year built 2012 Year refurbished Unknown Under cons. No Construction type Unknown Construction Unknown Façade description Type Unknown Unknown Cavity Unknown

Face Unknown material Core Unknown material Insulation Unknown material Cause Spark in ventilation system from cooking fat deposits Cause Electric_3

Starting floor 1 Starting floor 1 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Unknown Wind No_wind_effect

Falling material as a factor Unknown Falling material No

Sources 1. en.crimemoldova.com/news/social/another-fire-in-the-capital-the-roof-of-a-business-center-is- burning-/ 2. trm.md/en/social/incalcarea-normelor-antiincendiare-cauza-incendiului-de-pe-strada-ismail/

3. www.timpul.md/en/articol/Violation-of-anti-fire-rules-led-to-fire-in-shopping-center-on-Ismail- St-86850.html

98

Name Zen Tower, Dubai ID 59 Coordinates (Longitude, latitude, elevation above sea level) (25.0677870, 55.1308860, 11) Incident description The fire started at a low level on the outside of the building. Flames rapidly spread up the building and although there was a strong wind associated with a sandstorm there have not been reports that the wind accerated the fire. Reports suggest the building was newly constructed.

Primary database content Machine learning data Date 2018-05-13 Age 2018 Floors 15 Floors 15 Type Residential Type Residential Height (m) Unknown Height Unknown Country UAE Country UAE Year built approx. 2017 Year built 2017 Year refurbished n/a Under cons. No Construction type Unknown Construction Unknown Façade description Type Unknown Unknown Cavity Unknown

Face Unknown material Core Unknown material Insulation Unknown material Cause Unknown Cause Unknown_0

Starting floor Lower level Starting floor 1 Casulties None Fatalaties 0 Injuries 0 Sprinklers Sprinkler No_sprk_system Unknown

Wind as a factor Strong wind during sandstorm Wind No_wind_effect

Falling material as a factor Damge to cars Falling material Caused_property_damage

Sources 1. sputniknews.com/middleeast/201805131064407568-uae-dubai-fire-skyscraper/

2. https://gulfnews.com/news/uae/emergencies/residents-escape-dubai-tower-fire-1.2220689

99

100

APPENDIX C : MODEL EVALUATED ACCURACY

A series of variants of the example model descripted in Section 3.3 are trained with ranges of defined parameters in order to identify the optimal model parameters: • Training steps: 500, 1000, 2000 and 4000; • Trained samples: 6, 18, 30, 42; • Number of hidden layers: 1, 2, 3, 4; and • Number of neurons per hidden layer: 2, 3, 4, 5

Trained samples are randomly selected from the entire set of surveyed incidents where the unselected incidents for training purpose are used for testing the model accuracy. Given the limited database size, it becomes a challenge to cope with both training and testing the model. A reasonably greater number of training data is preferred to guarantee the data is relatively representative of a full spectrum of the space domain in each dimension and to limit overfitting. However, the testing data is proportionally reduced to the training data. Too few testing data renders the evaluated model inaccurate, for example, the model will be either 0 % or 100 % accurate if only one incident is supplied for testing. Therefore, a model with each set of parameters is trained 20 times (i.e. each accuracy in the tables are averaged over 20 trained models), and the final model accuracy is adopted as the average value of each individual model accuracy. 16 tables are shown in this appendix with each table containing 16 accuracies. Each table is a unique combination of number of hidden layers and number of neurons per layer, this is displayed at the top two rows of each table. Rows and columns show the total training steps and number of incidents that are used for training. The tables are also colour coded in line with their values, red and blue represents the higher and lower model accuracy, respectively. neurons_per_layer 2 layers 1

Sum of accuracy Training incidents Simulation steps 6 18 30 42 500 48% 54% 58% 64% 1000 50% 48% 54% 62% 2000 47% 47% 50% 55% 4000 46% 51% 51% 49%

101

neurons_per_layer 3 layers 1

Sum of accuracy Training incidents Simulation steps 6 18 30 42 500 50% 53% 54% 59% 1000 40% 48% 51% 56% 2000 42% 44% 53% 55% 4000 44% 46% 48% 47% neurons_per_layer 4 layers 1

Sum of accuracy Training incidents Simulation steps 6 18 30 42 500 45% 47% 55% 56% 1000 48% 49% 50% 54% 2000 44% 45% 50% 51% 4000 42% 49% 44% 45% neurons_per_layer 5 layers 1

Sum of accuracy Training incidents Simulation steps 6 18 30 42 500 49% 49% 55% 54% 1000 41% 47% 53% 50% 2000 47% 47% 51% 48% 4000 40% 46% 46% 48% neurons_per_layer 2 layers 2

Sum of accuracy Training incidents Simulation steps 6 18 30 42 500 51% 57% 57% 57% 1000 55% 57% 53% 57% 2000 47% 54% 50% 53% 4000 42% 54% 49% 57%

102

neurons_per_layer 3 layers 2

Sum of accuracy Training incidents Simulation steps 6 18 30 42 500 45% 49% 57% 61% 1000 46% 52% 51% 53% 2000 48% 44% 49% 49% 4000 46% 50% 47% 53% neurons_per_layer 4 layers 2

Sum of accuracy Training incidents Simulation steps 6 18 30 42 500 42% 51% 55% 51% 1000 48% 45% 51% 51% 2000 44% 44% 46% 47% 4000 39% 48% 42% 44% neurons_per_layer 5 layers 2

Sum of accuracy Training incidents Simulation steps 6 18 30 42 500 47% 47% 47% 53% 1000 45% 47% 47% 49% 2000 43% 45% 47% 47% 4000 43% 44% 46% 42% neurons_per_layer 2 layers 3

Sum of accuracy Training incidents Simulation steps 6 18 30 42 500 52% 56% 58% 58% 1000 49% 59% 60% 58% 2000 52% 58% 57% 56% 4000 56% 55% 54% 53%

103

neurons_per_layer 3 layers 3

Sum of accuracy Training incidents Simulation steps 6 18 30 42 500 50% 53% 60% 63% 1000 49% 52% 53% 56% 2000 45% 48% 49% 53% 4000 43% 50% 49% 50% neurons_per_layer 4 layers 3

Sum of accuracy Training incidents Simulation steps 6 18 30 42 500 48% 52% 54% 56% 1000 51% 41% 51% 50% 2000 38% 46% 43% 52% 4000 38% 42% 43% 49% neurons_per_layer 5 layers 3

Sum of accuracy Training incidents Simulation steps 6 18 30 42 500 45% 43% 52% 53% 1000 50% 40% 47% 47% 2000 37% 45% 41% 44% 4000 43% 42% 40% 39% neurons_per_layer 2 layers 4

Sum of accuracy Training incidents Simulation steps 6 18 30 42 500 54% 59% 59% 61% 1000 55% 59% 59% 57% 2000 59% 61% 60% 60% 4000 57% 56% 58% 58%

104

neurons_per_layer 3 layers 4

Sum of accuracy Training incidents Simulation steps 6 18 30 42 500 52% 56% 56% 61% 1000 50% 53% 55% 56% 2000 53% 51% 51% 55% 4000 38% 50% 51% 48% neurons_per_layer 4 layers 4

Sum of accuracy Training incidents Simulation steps 6 18 30 42 500 38% 47% 50% 51% 1000 53% 48% 49% 53% 2000 47% 43% 46% 49% 4000 40% 40% 42% 44% neurons_per_layer 5 layers 4

Sum of accuracy Training incidents Simulation steps 6 18 30 42 500 42% 46% 49% 50% 1000 41% 40% 47% 48% 2000 43% 41% 44% 46% 4000 37% 41% 39% 40%

105

APPENDIX D: DETAILED TRAINED MODEL INFORMATION

Data fields and associated feature column properties

Data fields and associated feature column properties

Feature key Description Data type Feature type Boundaries or vocabulary list

floors Number of int Bucketised [0, 15, 30, 45, 60] total storeys numeric

floors_numeric Number of int Numeric - total storeys

primary_type Primary use str Categorical ['commercial', 'hospital', 'hotel', of the 'office', 'residential'] building

country Country that str Categorical ['australia', 'azerbaijan', 'canada', the building 'chechnya', 'china', 'england', is located 'france', 'germany', 'hungary', 'indonesia', 'israel', 'japan', 'qatar', 'romania', 'russia', 'scotland', 'south korea', 'turkey', 'uae', 'uk', 'usa']

year_built Year the int Bucketized [0, 1960, 1970, 1980, 1990, 2000, building was numeric 2010] built

year_built_numeric Year the int Numeric - building was built

construction Building str Categorical ['-1', 'concrete', 'steel'] construction type

facade_type Building str Categorical ['-1', 'eifs', 'mcm'] façade type

106

cavity Is there any str Categorical ['-1', 'none', 'ventilated'] cavity in façade

face_material Building str Categorical ['-1', 'aluminium', 'mineral fibre surface board', 'polymer', 'render', 'steel', material 'zinc']

core_material Building core str Categorical ['-1', 'pe', 'polymer'] material

insulation_material Building str Categorical ['-1', 'combustible', 'eps', 'foam', insulation 'mineral wool', 'pir', 'pu'] material

sprinklers Sprinkler str Categorical ['no_sprk_system', system 'sprk_contained_2', characteristi 'sprk_not_activated', cs 'sprk_not_known_0']

under_construction Is the str Categorical ['no', 'yes'] building under construction

APPENDIX E: PREDICTED SEVERITY CLASSIFICATION SCORES OF THE BEST-PERFORMING OPTIMISED MODEL

Prediction probability Predicted Actual Data index [%] classification classification

1 74.8 0 3

2 66.2 0 1

3 66.2 0 0

4 80.9 0 0

5 99.5 1 1

6 84.7 0 7

7 82.8 0 0

107

8 75.2 0 3

9 64.3 0 1

10 81.7 0 0

11 70.1 0 2

12 99.6 7 7

13 89.1 0 0

14 99.7 9 9

15 99.6 0 0

16 82.8 0 0

17 84.3 0 0

18 81.9 0 0

19 65.5 7 7

20 88.9 0 0

21 97.1 8 8

22 87.2 7 7

23 99.6 0 0

24 72.7 0 0

25 69.2 7 7

26 81.8 0 0

27 73.4 0 0

28 73.4 0 0

29 67.6 0 0

30 99.9 0 0

31 88 0 0

32 89.6 0 0

33 81 0 0

108

34 89.1 0 0

35 81.2 0 0

36 99.8 1 1

37 99.8 0 0

38 83.3 0 11

39 99.6 4 4

40 81.5 0 1

41 80.5 0 2

42 73.3 0 1

43 85.9 0 0

44 88.2 0 0

45 80.6 0 3

46 99.5 0 0

47 86.6 0 0

48 99.6 0 0

49 90.1 0 0

50 79.6 0 0

51 86 0 0

52 90.9 0 0

53 83.9 0 0

54 99.2 0 0

55 54.7 0 7

56 84.5 0 0

57 79.6 0 10

58 86.9 0 0

109