Author(s): Serdar Selamet, Liyu Wang, Samson Min Cong Chau, and Sergio Chimal Ramirez
Abstract: Accurately predicting structural response under fire conditions is crucial for the safety and resilience of steel structures. Performance-based fire engineering often relies on finite element analysis, which can be time-consuming and computationally expensive. This study explores a more efficient approach to fire protection by using machine learning models to predict structural responses. Using data from parametric studies performed with OpenSees, we trained machine learning models, such as Artificial Neural Networks (ANNs) and Random Forests, to predict key structural metrics, including displacement, rotation, and demand-to-capacity ratio. The models consider input features like beam length, loading, and boundary conditions, and their performance is evaluated using metrics such as ROC-AUC for classification and R²-score for regression. Results show that these models approximate finite element simulations with a high degree of accuracy. This approach provides engineers with a faster way to estimate structural behavior under thermal loading conditions, reducing the need for extensive simulations and supporting more efficient, performance-based fire protection strategies.
Author(s): Ramin Yarmohammadian, Balsa Jovanovic, Ruben Van Coile
Abstract: The computational expense of advanced numerical methods in fire safety engineering often limits their practical application. This study introduces physics-informed surrogate models to efficiently predict the maximum temperatures of fire-protected steel elements exposed to natural fire scenarios. This offers a balance between computational speed and physical reliability. A numerical model integrating the Eurocode parametric fire curve (EPFC) and the lumped mass approach was developed to generate a dataset of 3,000 fire scenarios. Key input features, such as compartment dimensions, fire load densities, and insulation properties, were selected based on their physical relevance to heat transfer. An independent test set of 105 scenarios validated the models' performance. Initial black-box models, including regularised linear regression and neural networks, demonstrated strong predictive accuracy (R2 ≈ 0.99 for neural networks) but exhibited physical inconsistencies, such as predicting steel temperatures exceeding fire temperatures. These limitations highlighted the need for physics-informed enhancements. Key improvements included physics-guided feature selection, ensuring inputs aligned with thermal behaviour, and embedding physical constraints via custom loss and regression functions to prevent physically implausible predictions. Weighted training prioritised accuracy in critical high-temperature ranges, while sub-models captured the distinct behaviours in fuel- and ventilation-controlled fires. Sensitivity analysis underscored the significance of factors such as maximum temperature in the fire and insulation thickness. The results demonstrate the potential of physics-informed surrogate models to deliver rapid and reliable temperature predictions. These advancements address critical challenges in performance-based fire safety design and pave the way for real-time risk assessment tools in structural fire engineering.
Author(s): M. Hamed Mozaffari, Yuchuan Li, Yoon Ko
Abstract: NRC Fire Safety VISTA (vision, internet of things, safety, technology, artificial Intelligence) laboratory is at the forefront in developing cutting-edge technologies, methodologies, and algorithms utilizing digital technologies, AI, IoT, and CV to enhance fire safety for Canadians. The NRC research activities have explored the potentials of digital technologies, Artificial Intelligence (AI), Computer Vision (CV), and the Internet of Things (IoT), robotics, remote sensing, and UAVs, in application to the various fire safety engineering areas, including fire modelling; fire detection; IoT monitoring; firefighting, and fire investigation as well as fire risk assessment. To advance detection technologies, NRC VISTA devices were utilized to develop algorithms for smoke and flame pattern analyses based on thermal and vision data, which resulted in an innovate gas leak detection technology, and vision-based fire detection and flashover prediction technologies. For firefighting assistance, AI tools were developed to assist first responders in HAZMAT fire identification based on combustion signature analysis and decision-making for emergency responses by providing corresponding emergency response strategies in real-time (supported by NRC AI4Logistics program). Also, a first responder training tool is also developed based on a virtual reality technology for transportation emergency and management. Our work also addresses fire data management, which can assist fire investigation and fire risk assessment. An AI based tool to support fire investigation is developed for decoding particle images of soot agglomerated in smoke detectors. NRC team is also leading the integration of fire risk assessment into digital building information management platforms, which enable to leverage AI, CV, and IoT for automation, real-time monitoring and instant alerts to homeowners and emergency services, thus significantly reducing response times and potential damage.
Author(s): Yuchen Wang, Mhd Anwar Orabi, Zhuojun Nan, Wei Ji, Xinyan Huang, Asif Usmani
Abstract: Building fires occur frequently and pose significant hazards, making fire safety design critically important. The application of fire protection serves as an effective technique to maintain structural stability of a burning building. However, fire protection coatings contribute substantially to the overall cost of a steel building construction, highlighting the need for efficient and optimized design strategies. Moreover, current prescriptive approaches often result in suboptimal solutions that may fail to ensure the required level of fire resistance, as they are based on idealized fire scenarios that ignore nonuniform temperature distribution. In this paper, we explore the potential of AI-driven design for structural fire protection solutions under more realistic fire conditions. This approach first derives an optimized fire protection layout by simulating various localized fire scenarios on a real structure. The simulation results are used to evaluate individual structural components by comparing them with their pre-determined critical temperatures, allowing the identification of critical protection zones. A Graph Neural Network (GNN) model is then trained on the numerical simulation dataset. Its mesh-based representation, consisting of elements and nodes, allows it to effectively capture both local and global structural features. This enables the optimized design approach for fire protection layout to be transferable to similar structures. The GNN-based transfer learning method yields a fire protection design that significantly outperforms traditional approaches, demonstrating enhanced efficiency, cost-effectiveness, and adaptability under realistic fire conditions.
Author(s): Sai Pavan Kumar Balabomma, Bogdan Matuszewski, Eleni Asimakopoulou
Abstract: The increasing focus on performance-based fire safety regulations highlights the need for specialised simulation tools for fire risk assessment, evacuation strategies, and structural safety. While Computational Fluid Dynamics (CFD) and Finite Element Analysis (FEM) tools provide detailed insights, they are resource-intensive and have limitations. Advanced deep learning (DL) methods subset of machine learning (ML) show significant promise in identifying complex patterns within data. This study introduces a novel Recurrent Neural Network RNN surrogate model designed to predict cladding surface temperatures during BS8414 façade fire tests, a critical scenario in fire safety engineering.
Author(s): Yiming Jia and Eyitayo A. Opabola
Abstract: Wildfires represent an escalating threat to California’s buildings, infrastructures, and communities. Especially, the January 2025 Southern California wildfires have underscored the urgent need for predictive tools that not only assess wildfire-induced damage but also can be integrated into wildfire risk models. This study introduces an interpretable machine learning framework developed to serve as a damage modeling tool within wildfire risk assessments. Using the CAL FIRE damage inspection dataset, structural damages are categorized into two distinct classes, unaffected-to-minor and moderate-to-severe, based on a threshold defined as 10% of the structure’s value. A set of 21 features are derived from the fields reported in CAL FIRE damage inspection dataset and categorized into three risk dimensions: hazard, vulnerability, and exposure. A comprehensive suite of candidate models is considered, including both machine learning and statistical models: deep neural network, symbolic neural network, random forest, balanced random forest, XGBoost, support vector machine, k-nearest neighbor, and logistic regression. Pre-2025 wildfire data is leveraged to predict the damage state of structures affected by the January 2025 Southern California wildfires, providing a real-world case study to evaluate model performance. Among the candidates, the deep neural network outperformed the other models, demonstrating superior predictive accuracy and an exceptional ability to capture the complex, nonlinear relationships between the input features and the observed damage states. A subsequent 5-fold cross-validation is applied on the entire CAL FIRE damage inspection dataset to refine the deep neural network, which ensures that its generalizability and robustness under diverse wildfire scenarios. Importantly, this refined deep neural network is intended to be incorporated into wildfire risk models, offering end users (e.g., engineers, stakeholders, and decision-makers) a practical tool for resilience assessment and mitigation strategy development. To enhance model transparency and interpretability, the Shapley Additive Explanations (SHAP) method is employed. SHAP provides both local and global interpretations for the deep neural network. At the local level, SHAP quantifies feature impact by illustrating how each feature affects the damage state prediction for an individual record. At the global level, it evaluates overall feature importance across the entire dataset, quantifying the contribution of each feature to wildfire damage state estimation. These interpretative insights are essential for understanding the key drivers of wildfire damage, which provides actionable guidance for preparing for wildfire hazards, including prioritizing structural improvements, optimizing resource allocation, and developing more targeted emergency response strategies. Overall, this research bridges advanced machine learning techniques with the practical demands of wildfire risk modeling. The resulting deep neural network model, coupled with detailed SHAP-based analysis, offers a practical and robust tool for future wildfire damage prediction and can be seamlessly incorporated into wildfire risk models, thereby contributing to enhanced resilience and improved hazard preparedness in California.
Author(s): Yizhou Li, Yanfu Zeng, Xinyan Huang
Abstract: The increasing frequency of wildfires, particularly in the urban-rural interface zones of Asia, underscores the urgent need for advanced real-time simulation and forecasting models to mitigate the risks to human life and property. This study introduces a novel computer vision-based operational model for wildfire management, specifically designed for the wildland-urban interface of Hong Kong Island. Utilizing a dataset of 240 cases (8640 samples) derived from FLAMMAP simulations, we develop a deep learning framework capable of modeling wildfire dynamics at a high spatial resolution of 5 meters, based on satellite imagery. The model tackles the challenge of training on fine-resolution satellite data, which is complicated by the complex and heterogeneous nature of wildfire spread. Our methodology incorporates a multi-stage approach to model the evolution of wildfires, distinguishing between early-stage and large-scale wildfire behaviors, and facilitating effective cross-scale segmentation. This segmentation strategy prioritizes the core burning regions, which are essential for accurate real-time forecasting. Evaluation results show that our deep learning model achieves an accuracy of 88.6% in predicting wildfire spread, offering valuable operational insights for real-time simulation and disaster management in Hong Kong. The proposed model demonstrates significant potential for enhancing wildfire management efforts, enabling the generation of rapid, high-resolution forecasts that can be integrated into cloud-based wildfire monitoring systems.
Author(s): Nikolaos Kalogeropoulos, Guillermo Rein
Abstract: Wildfires are becoming bigger and faster, affecting more areas and causing more fatalities and damage. Studying wildfire behaviour through computer modelling is required to understand and plan against future wildfires. These models use assumptions and region-specific constants, yielding different and sometimes contradictory results. In this paper we compare four state-of-the-art wildfire models along with two machine-learning surrogates to study their differences. Of the models selected for this study, Farsite, Prometheus, ELMFIRE and FDS Level Set have been developed to simulate wildfire spread in North America, while Google’s EPD and EPD-ConvLSTM have been developed as machine learning surrogate models of Farsite. The benchmarking is done through simple scenarios and a real wildfire, Mati 2018 in Greece, chosen because it is outside all the models’ calibration regions. Compared to Farsite, ELMFIRE and Prometheus show 40% slower flank- and back-fire, with the same head fire and produce smaller wildfire scars. Prometheus has the highest similarity of the final scar to the real fire, 59%. The Google models do not provide realistic results in the simple scenarios but produce similar real wildfire scars to Farsite. FDS Level Set predicts 50% larger fire scars. This paper shows, through a comparison of the six models, that ensemble modelling is beneficial for a prediction of wildfire behaviour. Using ensemble modelling results in a probabilistic map of wildfire spread, where differences and agreements can be studied, providing a better understanding of wildfire extent.
Author(s): Mohammad al-Bashiti, Arash Teymori, Zoie Mccreery, M.Z. Naser
Abstract: Fire-induced spalling of concrete remains a challenging and complex phenomenon. A thorough review of current literature reveals significant inconsistencies among prevailing theories and underscores the difficulties in accurately predicting concrete spalling under elevated temperatures. In response, we introduce an innovative approach that harnesses the latest advances in explainable Artificial Intelligence (XAI) to create a heuristic model for predicting spalling in concrete mixtures. This study presents our development and validation of an XAI model, enhanced by Shapley Additive Explanations (SHAP), based on data from over 1,000 fire tests. The proposed XAI model not only can predict the fire-induced spalling with high accuracy (i.e., >94 %) but can also articulate the reasoning behind its predictions (as in, the proposed model can specify the rationale for each prediction instance); thus, providing us with valuable insights into the factors, as well as relationships between these factors, leading to spalling. The analysis identifies eight primary factors that govern spalling behavior: the presence of polypropylene fibers, moisture content, heating rate, maximum exposure temperature, silica fume/binder ratio, sand/binder ratio, water/binder ratio, and fly ash/binder ratio. Although these factors have been recognized in existing theories, our work quantifies their contributions, for the first time, to reveal a significant variability in their impact. Additionally, the study introduces a practical nomogram that enables researchers and engineers to visually assess the spalling propensity of any given concrete mixture with ease.
Author(s): Yifei Ding, Xinghao Chen, Yuxin Zhang, Xinyan Huang
Abstract: The real-time crowd data extraction from surveillance devices is essential for the onsite emergency decision-making. However, traditional evacuation monitoring system using single-camera tracking is limited to interoperating data from distinct cameras leading to redundant and inaccurate information, which results in huge necessity for multi-camera tracking technology. In this work, a novel real-time multi-camera tracking framework using an improved re-identification (Re-ID) model is proposed for evacuation safety monitoring. The framework consists of (1) multi-camera network, (2) human detection model, (3) tracking model, (4) an explainable attention-aided Re-ID model, and (5) module of feature matching and re-distribution algorithm. It enables the detection, tracking and re-identification of evacuees across multi-camera. Furthermore, to demonstrate real-time multi-camera tracking, a simplified evacuation drill is conducted to demonstrate real-time multi-camera tracking, showing good accuracy in Re-ID and personnel counting, where the overall Re-ID tracking accuracy exceeds 75% and the personnel counting accuracy is around 100%. Lastly, the class activation map (CAM) illustrates the model explainability and limitations. Overall, our approach enables the tracking and re-identification of individuals across different camera views which significantly enhances surveillance capabilities and contributes to the development of more automatic and intelligent monitoring systems.
Author(s): Scott Silverstein
Abstract: This project aims to develop an automated system using Computer Vision (CV) to process ground-level and aerial imagery of Wildland-Urban Interface (WUI) properties. The goal is to identify key risk indicators and generate structured data on hazards and mitigation needs, thereby augmenting traditional manual inspections and enabling broader, more consistent data collection. The system employs CV models (evaluating DETR/Yolo variants for detection, SAM for segmentation) to analyze images, automatically identifying and classifying features like roof materials, vents, vegetation proximity/condition, and defensible space characteristics. The output is structured, property-specific risk data (e.g., flammable roof, hazardous vegetation) and potential mitigation suggestions. The structured data serves multiple purposes for stakeholders: (1) Trend Analysis: Enables local governments to identify community-wide risk patterns for targeted mitigation planning; (2) Digital Twin Foundation: Provides data for creating digital property/neighborhood models for planning and simulation; (3) Stakeholder Information: Offers clear risk/mitigation data for property owners and potential insights for insurance risk assessment. Research focuses on evaluating state-of-the-art CV models (DETR, YOLO, SAM, etc.) to maximize accuracy and reliability in feature detection and segmentation, prioritizing robustness over real-time speed. The project seeks to deliver a system generating consistent, scalable WUI risk data from imagery. This automated data collection and analysis will support property owners, local governments, and insurers in making informed mitigation decisions and complementing existing efforts as envisioned by the SFPE GCI.
Author(s): Yanfu Zeng, Xinyi Liu, Yifei Ding, Tianhang Zhang, Xinyan Huang, Xinzheng Lu
Abstract: Buildings have been carefully designed in terms of fire safety to safeguard the lives and properties under the occurrence of fire incidents. However, current fire safety design process suffers from time-consuming and computational-expensive analysis of fire-smoke behaviours, as well as intensive workloads and inevitable human errors due to manual preparation of the engineering drawings. This poster presents the latest research advancements in AI applications for digitalizing fire safety design, addressing these challenges through deep learning-powered solutions. The first major advancement is the development of an AI-powered tool, the Intelligent Fire Engineering Tool (IFETool), designed to accelerate fire safety assessments. Using a transposed convolutional neural network (TCNN) trained on a database of over 1,000 fire simulations, the tool can predict critical fire safety parameters, such as smoke visibility, gas temperature, and CO concentration. These predictions, which traditionally require up to two days of numerical modelling, can now be obtained in one second with 97% accuracy. The tool also features an open-access interface (http://ifetool.firelabxy.com/) that provides essential design information, including available safe egress time (ASET) and smoke evolution curves, allowing engineers to assess fire safety rapidly and identify design limitations. To expand its applicability to complex architectural structures, authors further explore the use of Generative Adversarial Networks (GANs) to predict fire-smoke behaviour in buildings with varied floor plans and fire locations. A GAN-based model trained on 136 fire simulations successfully captures spatial patterns of fire behaviour, including plume formation, heat distribution, and smoke movement, with an 88% accuracy. Additionally, the AI model estimates activation times for heat detectors and sprinklers with a 95% precision, providing a fast and reliable alternative to Computational Fluid Dynamics (CFD) simulations, which are computationally intensive and costly. Further advancements extend the AI framework to three-dimensional fire behaviour modelling, addressing the needs of modern architectural designs. By employing GANs and diffusion models, this study predicts smoke propagation in atriums and other complex spaces within seconds, achieving an accuracy of up to 92%. The AI-generated smoke visibility data can be used to estimate safe egress times more efficiently than traditional methods. While the diffusion model provides detailed local smoke patterns, its longer rendering time (20 minutes) makes it suitable for high-accuracy design assessments. Beyond fire-smoke analysis, this research explores automated sprinkler layout design using deep learning. A pix2pixHD GAN model, trained on 120 sprinkler design drawings, generates code-compliant sprinkler layouts within seconds, achieving over 99% coverage. This AI-driven approach drastically reduces the manual workload of fire engineers, allowing them to focus on higher-level design and optimization tasks rather than repetitive drafting work. The findings presented align with the research priorities outlined in the SFPE Foundation’s Grand Challenges Initiative Digitalization, AI, & Cybersecurity white paper, particularly in the domain of Digitalized DIOM of Fire Protection (FP) Systems. By integrating AI into fire safety design, this project contributes to the evolution of intelligent, data-driven fire protection solutions that improve accuracy, speed up decision-making, and foster innovation in the design process.
Author(s): Maryam Zamanialaei; Michael Gollner, Maria Theodori, Dwi Purnomo, Ali Tohidi, Chris Lautenberger, Arnaud Trouve, Yirin Qin, Daniel San Martin
The destructive impacts of Wildland-Urban Interface (WUI) fires on people, property, and the environment have dramatically increased, particularly in California, where extreme fire events have caused unprecedented losses. As WUI fires become more frequent and severe due to climate change and expanding urban development, understanding the factors that contribute to structure loss is critical for developing effective mitigation strategies. While several key factors influencing structure protection during wildfires—such as home hardening (e.g., vents, siding, roofing materials, eaves, window type, and construction year), defensible space, exposure to flames and embers, and structure separation—are well recognized, their interrelated impacts on structure survival have not been fully quantified. In this study, we analyze five major historical WUI fires—the 2017 Tubbs, 2017 Thomas, 2018 Camp, 2019 Kincade, and 2020 Glass Fires—to assess patterns of structure loss and the effectiveness of mitigation measures. We integrate machine learning models with exposure data, offering a comprehensive, data-driven assessment of structural vulnerability in WUI fires. To enhance the accuracy of exposure quantification, we incorporate the outputs of a WUI fire spread model, allowing us to simulate flame impingement, ember transport, and structure-to-structure fire spread, enabling a more detailed evaluation of exposure pathways. To validate our exposure modeling, we link flame length and ember loads with damage inspection data (DINS) collected during post-fire assessments in California. The DINS dataset provides ground-truth information on structure damage classifications, allowing us to analyze how flames and embers directly contribute to observed structure loss patterns. By integrating historical fire case studies with fire behavior modeling, we develop a more precise understanding of the thresholds of exposure that lead to structure loss, which is crucial for refining wildfire risk assessments and mitigation strategies. We employ advanced machine learning models, including XGBoost and Random Forest classifiers, to predict structure survival probabilities based on a combination of exposure metrics and mitigation factors. This method significantly improves predictive accuracy, achieving an 82% precision rate in identifying at-risk structures. Our results demonstrate that home hardening and defensible space, particularly in the immediate zone around a structure (Zone 0), remain among the most effective mitigation measures. Our hypothetical loss reduction analysis shows that enhancing home hardening and vegetation clearance could reduce structure losses by up to 52%. Our findings highlight that structure separation and exposure to flames and embers are among the most significant factors influencing the probability of structure loss. These findings provide data-driven, actionable insights for policymakers, emergency planners, and homeowners, emphasizing the need for a multi-scale approach to fire mitigation in the WUI.
Author(s): Yanglan Wang, Kenichi Soga
Efficient traffic flow optimization is vital for effective traffic management, especially during emergencies such as evacuations, which require the rapid and safe movement of large numbers of vehicles under intense time constraints. This study presents an optimal control model designed to regulate traffic input over time in rural road networks, with the objective of minimizing total travel time. The model integrates link-specific dynamic equations, variable constraints, and the flow–density relationship. A real-world case study of an evacuation in Inverness, California, is used to demonstrate the model’s applicability. Additionally, a microscopic simulation is also conducted to validate that similar input in both microscopic and macroscopic simulations produce consistent output patterns. The results show that the optimal control model substantially reduces overall travel time compared to scenarios employing uniform input distribution.
Author(s): Yu Liu
Abstract: The growing complexity of fire protection regulations and technical documents necessitates efficient methods for accurate analysis and validation. This research explores AI-assisted document parsing combined with cross-validation and verification techniques to enhance the reliability of fire safety assessments. By leveraging natural language processing (NLP) and machine learning, AI can systematically extract, analyze, and interpret critical information from extensive documentation, reducing human error and improving efficiency. Cross-validation and verification methods further ensure data integrity by comparing extracted information across multiple sources, identifying inconsistencies, and enhancing decision-making confidence. This study presents the effectiveness in processing fire test report key value extraction, validation across documents, and result consistency verification throughout the report. The results highlight AI’s capability to improve accuracy and efficiency in fire protection engineering.
Author(s): Jakub Bielawski, Dia Luan, Xinyan Huang, Wojciech Węgrzyński
Abstract: Tunnel facilities enable transportation, including goods and people, supporting communication and logistics. Due to the significant role they play in public life and the possibility of exposing a large number of people to an emergency in it, it is essential to provide adequate safety level. Fires are one of the most common and potentially damaging events. The severity of fires and the risk associated with them, were observed in many catastrophic events. The current state-of-the-art in fire safety system design is based on a worst-case scenario approach. Considering the prescribed design fire models, there is a chosen maximum possible fire in terms of vehicle size and heat release rate. The most common smoke control solution for tunnels is mechanical ventilation. It is used whenever spatial arrangement and technical conditions do not allow the use of natural wind-driven ventilation. Basic and robust type of mechanical ventilation in longitudinal system based on jet fans placed in tunnel space. Design criteria such as critical ventilation velocity (CVV) are estimated using international standards. However, HRR evolution in actual fire events is unpredictable. It has been observed from literature data and our fire tests conducted in small-scale and actual tunnels that ventilation conditions significantly influence the overall tunnel fire environment. It affects diffusion flames' behaviour, energy balance, smoke flow direction and patterns, as well as thermal stratification and back-layering occurrence. An additional aspect is the initial conditions of airflow caused by atmospheric factors and vehicle traffic. Due to the unpredictable flow conditions during a real fire, it is impossible to provide adequate ventilation for every scenario with a single ventilation system configuration. A potential solution described in the summary is real-time dynamic adaptation to current conditions, monitored by a network of sensors. The main objective of my PhD study is to explore the extent to which it is possible to reasonably and dynamically control actual fire conditions in tunnels. It requires real-time sensor monitoring and adaptive models to predict incoming fire development. Investigation includes models like PID, fuzzy logic, genetic algorithms, and physics-based ML/AI. This kind of application using real-time sensor data is a part of a wide research field of experts at the AI Summit event. My intended outcome will consist of an overview of our recent studies of ventilation effects on fire safety systems, like fiber-optic linear heat detection, fire suppression systems as well as the ventilation effect of HRR and flame, energy balance in model tunnels. The object that narrows down the broad issue of variability of tunnel fire parameters will be the adaptive model, which has high applicability potential where advanced data processing supports fire safety engineering of actual tunnels.
Author(s): Jong Hoon Kim, Jung Hyun Yoo, Duck Hee Lee, Seok Woo Hong, Kyo Hyuk Lee
Abstract: This study was initiated with the purpose of developing a prototype system that predicts the fire behavior in deeper underground subway stations and tunnels, providing the information for firefighting, evacuation, and rescue operations. This system is being developed to provide the predictive information such as temperature and gas concentration from a fire to firefighters and emergency responders, based on the heat and smoke data measured at the fire scene. After designing the basic structure to predict and provide information based on the measured data from the scene, research on the prediction component was initiated. Initially, the chosen approach involves creating multiple fire scenarios for the target space. These cases are then analyzed through fire modeling. Subsequently, the results are built into a database. When a fire occurs, the system aims to pick up the case that has the closest predictive results based on incoming data such as temperature and gas concentration from the fire scene, and to predict the future situation following the time the data was received. To build the database for such a system, it is necessary to generate results from fire modeling that analyze various scenarios. To build the database, interpreting an immense number of scenarios requires significant cost and time, so we aimed to use an AI-based surrogate fire model for this part. While conducting research to generate additional fire modeling results using this, the participating researchers came to believe that creating a system for direct predictions using a trained AI-based surrogate fire model would be more efficient. Additionally, the research goal was shifted towards developing a system that quickly estimates the modeling results of fire phenomena based on real-time incoming fire data from the scene. In this process, existing research conducted by Yanfu Zeng et al. and Xiqiang Wu et al. from Hong Kong Polytechnic University was referred to. Thus, research was conducted to create an AI-based surrogate fire model using machine learning. Currently, this has been tested on the tunnel. A dataset of 108 fire scenarios targeting tunnel spaces was created, and the AI model was trained with this data. CFD images were compared with the predicted images from the AI-based surrogate model to assess accuracy. This research is also currently exploring the application of the AI-based surrogate fire model to underground stations.
Author(s): Karl Toepperwien
Abstract: Wildfire spread in complex terrain poses major threats for fire fighters as interactions between topography, wind, and fire-dynamics give rise to unexpected fire behavior. This study investigates the fire-line rotation in a canyon fire wherein the fire front progresses perpendicular to the nominal wind direction. Using large-eddy simulations with a physics-based mesoscale solver, we model coupled fire-atmosphere-terrain interactions over kilometre-scale domains to resolve the three-dimensional flow and combustion structures governing fire spread. By considering an idealized canyon terrain and comparing it against two simpler configurations, a sloped ramp and a flat surface, we show that, despite identical ridge slopes, the canyon induces distinctly different spread behavior, including oblique propagation angles and intermittent progression in the valley. A detailed examination of the plume structure and velocity field attributes these phenomena to terrain-induced wind misalignment and localized vorticity amplification, which persist after fire front passage and promote erratic spread. Furthermore, these results also show that the rate of spread in complex terrain is inherently non-local: individual sections of the fire line are influenced by neighboring segments, transient flow structures, and topographic features, making an a-priori estimation of local spread rates unreliable.
To accelerate simulations and enhance future integration with AI models, the solver (Swirl-FIRE) used in this work is implemented in TensorFlow and executed on Tensor Processing Units (TPUs) using just-in-time compilation. This architecture leverages the same ML hardware and software ecosystem used in AI research, enabling rapid, scalable simulations of large-scale wildfire scenarios while maintaining physical fidelity.