Machine Learning for Fire Safety in the Built Environment: A Bibliometric Insight into Research Trends and Key Methods

Abstract

Assessing building fire safety risks during the early design phase is vital for developing practical solutions to minimize loss of life and property. This study aims to identify research trends and provide a guiding framework for researchers by systematically reviewing the literature on integrating machine learning-based predictive methods into building fire safety design using bibliometric methods. This study evaluates machine learning applications in fire safety using a comprehensive approach that combines bibliometric and content analysis methods. For this purpose, as a result of the scan without any year limitation from the Web of Science Core Collection-Citation database, 250 publications, the first of which was published in 2001, and the number has increased since 2019, were reached, and sample analysis was performed. In order to evaluate the contribution of qualified publications to science more accurately, citation counts were analyzed using normalized citation counts that balanced differences in publication fields and publication years. Multiple regression analysis was applied to support this metric’s theoretical basis and determine the impact levels of variables affecting the metric’s value (such as total citation count, publication year, and number of articles). Thus, the statistical impact of factors influencing the formation of the normalized citation count was measured, and the validity of the approach used was tested. The research categories included evacuation and emergency management, fire detection, and early warning systems, fire dynamics and spread prediction, fire load, and material risk analysis, intelligent systems and cyber security, fire prediction, and risk assessment. Convolutional neural networks, artificial neural networks, support vector machines, deep neural networks, you only look once, deep learning, and decision trees were prominent as machine learning categories. As a result, detailed literature was presented to define the academic publication profile of the research area, determine research fronts, detect emerging trends, and reveal sub-themes.

1. Introduction

The study of fire safety in the built environment has taken on different dimensions as technology has developed. Historically, there have been three fundamental approaches in the field of fire safety: prescriptive analysis (PA), performance-based analysis (PB), and artificial intelligence-based analysis (AI). Understanding the contributions of these approaches to the design process is critical, particularly in determining the future position of machine learning-based methods [1,2]. The study of fire safety in buildings started with written legislation, which is clear, binding, and mandatory. However, the adaptation of prescriptive legislation to fire risks is insufficient due to the development of construction systems, material technologies, and changing climate conditions. In addition, the need to incorporate PA into design has led to the problem of design constraints [1]. In this case, performance-based fire safety analysis has emerged with the capabilities of computer technology. PB analysis is a fire safety approach focusing on meeting specific safety objectives rather than fixed regulatory rules. It is carried out by considering the structure’s physical characteristics, user behavior, and possible fire scenarios. This method aims to provide more flexible, situation-specific, and applicable solutions through engineering calculations and simulation-supported analyses. As a result, more flexible, applicable, and realistic safety solutions based on experiments and/or simulations have emerged, which can go beyond the limits imposed by regulations and minimize unforeseen situations by addressing design-specific risks. PB analyses enable designs to be liberated and fire risks to be implemented and monitored during the design phase. Thus, fire safety designs based on experiments and/or simulations that can go beyond the limits given by the provisions and minimize unforeseen situations by addressing design-specific risks have emerged [3,4,5,6]. In recent years, innovative AI approaches have increasingly been incorporated into the design, evaluation, and prediction processes in the built environment [7,8,9]. Although PB approaches offer flexible and realistic solutions during the design phase, the length of the model and simulation periods limits dynamic data flow and updates. Predictive fire safety is achieved through the use of machine learning algorithms. Instead of fixed rules or static simulations, it makes predictions with real-time data flow and automates decision mechanisms. Unlike traditional methods, machine learning can discover multivariate, non-linear relationships based on data, enabling new insights and predictability in fire safety design by modeling complex parameters such as previously undefined fire scenarios, risk distributions, and user behaviors, thereby offering new insights and predictability in fire safety design [10,11,12,13]. A comparative table is provided to understand the contribution levels of the transition process from a rule-based approach to performance and AI-supported design (

Table 1. PA, PL, AI differences.

Features	PA	PB	AI
Flexibility	Low	Medium	High
Time	Low	High	Very low
Simulation	-	Existent	Existent
Real Time Data	-	Local	Existent
Preventive Measures	-	Medium	High
Predictive Safety	-	Medium	High
Sensing and Monitoring	-	-	High
Human Behavior	-	Existent	Existent

).

Since the prescriptive-based approaches created standardization in fire safety requirements and the designs could not be verified, a tendency towards performance-based design approaches for fire safety design began in the late 20th century. Including engineering tools and numerical analysis in fire science has provided important opportunities for the construction sector and scientific studies [14]. Countries such as America, Japan, England, and New Zealand are pioneers internationally as they integrate performance-based approaches into standards and codes for fire safety.

New Zealand was one of the first countries to work in this area by adding a section of the Building Act performance-based design (PBD) to the code-named New Zealand Building Code—Fire Safety (C/AS1), which was initially presented as prescriptive-based. However, the full integration of a performance-based system into the codes was achieved by publishing the C/VM2 (Verification Method for Fire Safety Design) document in 2001 [15,16].

In England, the first performance-based articles were introduced with the Approved Document B (ADB) standard, which was added to the prescriptive-based legislation implemented in the 1960s [17]. As of 2006, the ADB and BS 7974:2001 standards were updated to include performance-based solutions in large-scale and complex projects [18].

In the USA, in 1996, the Society of Fire Protection Engineers (SFPE) published the “Engineering Guide to Performance-Based Fire Protection Analysis and Design of Buildings” [19]. This guide sets out basic principles for evaluating the applicability of performance-based design. As of 2000, the National Fire Protection Association (NFPA) adopted a hybrid system that fully allows but does not mandate, instead using performance-based design through codes that were developed [20,21,22].

Japan added approaches that allow PBD solutions in fire safety design in the radical revisions made in 1998 to the “Building Standard Law of Japan (BSL)” standard, which came into force in 1950. In 2000, it harmonized PBD services with international standards by publishing “Fire Safety Engineering Guidelines” [22,23].

PBD approaches began developing in the 1990s through national building codes and academic research. Countries such as New Zealand, the USA, the UK, and Japan adopted PBD and created the first codes and standards in this field. Academic studies began to investigate PBD methods with numerical and experimental analyses as an alternative to prescriptive approaches [24,25]. Today, PBD continues to develop and provide flexible and reliable solutions by integrating artificial intelligence and advanced modeling techniques. The widespread impact and research trends of the use of artificial intelligence, which is the end of the three-stage process of fire safety design and provides predictive solutions to different problems in buildings in the field of fire safety, are examined in this study. Considering the dominant effect of predictive methods on designs for built environments, this study focuses on machine learning methods.

The objective of this study is to examine the current status, trends, and methodological diversity of machine learning-based research in fire safety through bibliometric analysis and qualitative content analysis. In this context, this study investigates which subtopics of fire safety have been the focus of artificial intelligence and machine learning-based modeling in recent years, which algorithms have been preferred, and what types of data have been used in these studies.

In this regard, the main research questions of this study are as follows:

• Which countries, authors, journals, and publications have primarily integrated machine learning into fire safety problems?
• In which thematic categories is machine learning applied within fire safety?
• What are the most frequently used machine learning algorithms, and what are the reasons behind these preferences?

2. Methodology

In order to reflect the trends in literature research, it is important to analyze the progress and acceleration of studies over time. Such analyses help us to understand the dynamics in research areas and the rise of specific topics in the literature [26,27]. Since this study uses bibliometric and qualitative content analysis, maintaining data integrity and methodological consistency has been a priority. In this regard, preliminary searches conducted in the Web of Science (WoS) and Scopus databases revealed that WoS provides more qualitative results regarding publication, citation counts, and publication quality. The WoS database was chosen due to its consistency in data standardization and extensive indexing capacity [28]. In January 2025, a comprehensive search was performed using the search query in

Table 2. WoS keywords research.

Topic		“fire” or “smoke” or “fire detection” or “evacuation “
And	Topic	“built environment” or “building” or “civil structure”
And	Topic	“machine learning” or “deep learning” or “neural network” or “predictive modeling” or “decision tree” or “vector machine”
Not	Topic	“forest” or “wildfire” or “medical” or “car” or “road”

, and 250 articles were identified. The WoS search covered the title, abstract, and author keywords fields. The search strategy was structured around three main categories using WoS’s ‘AND’ operator: ‘fire,’ ‘built environment,’ and ‘machine learning.’ To exclude studies not directly related to fire safety in buildings, exclusion operations were applied using the ‘NOT’ operator for the terms’ forest,’ ‘wildfire,’ ‘medical,’ ‘car,’ and ‘road.’ This method focused solely on publications related to building-based fire safety and machine learning. Although there was no year limitation in the WoS search process, studies in the research field started in 2001. As of 2019, the increase in the number of machine learning predictions included in fire safety design studies showed that predictive methods have become widespread in the research field in the last 5 years (Figure 1). The distribution of different types of publications is analyzed in Figure 2. The highest number of articles (191) indicates that the research is comprehensive. Conference papers (Proceeding Paper) are in second place with 55, indicating that it is an active research area for rapidly sharing innovative ideas. The low number of review articles indicates the lack of general systematic evaluations of the research area [29,30,31].

According to the analysis data obtained from the Web of Science category, there are 67 research field categories. However, 19 of these categories have one study, while the number of categories with at most five studies is 43. While the large number of categories indicates the interdisciplinary prevalence of the research field, the low number of studies in more than half provides an understanding that the research intensity covers specific disciplines. Figure 3 shows the categories with more than five studies according to the Web of Science category, and the most effective fields are Engineering Electrical Electronic (58), Engineering Civil (45), Computer Science Artificial Intelligence (34), Construction Building Technology (34), Engineering Multidisciplinary (34), and Materials Science Multidisciplinary (31).

In this study, multiple verification methods were used to increase the reliability of the findings. In the first stage, 250 publications on fire safety and machine learning selected from the Web of Science database were reviewed. Subsequently, 40 effective studies classified according to citation and normalized citation values were subjected to content analysis. Through content analyses, both the main themes and the prominent machine learning models were identified. The classifications obtained were compared with VOSviewer 1.6.20 analyses based on author keywords, and a high degree of similarity was observed between the two methods.

In evaluating the analysis data, normalized citation counts were used in addition to absolute metrics (number of publications and citations) to balance the effects of year and study category. Normalized citation reduced time and category-related imbalances, resulting in more balanced outcomes at the country, author, and publication field levels. All these methods supported the consistency of the trends and subheadings identified in this study, both qualitatively and quantitatively.

3. Bibliometric Analysis Results

This study’s bibliometric analysis examines research integrating fire safety and machine learning methods. Bibliometrics is a quantitative analysis system that measures numerical and statistical data in a specific research area [27]. Bibliometric analysis, which provides visually powerful materials for analyzing research, identifies research, authors, countries, publishers, and categories in a specific area [32]. Bibliometric analysis provides comprehensive data on a specific field by processing the literature. A bibliometric study, on the other hand, provides a broad perspective on the literature from which comprehensive data are obtained and allows quantitative and objective definitions of research topics from the past to the present [33]. This study presents a comprehensive review of predictive methods in fire safety design using bibliometric analysis. It aims to identify the areas of application of machine learning methods, which have become a widespread area of research in recent years, in the field of fire safety and the subcategories of machine learning.

“VOSviewer,” developed by Van Eck and Waltman in 2010 at the Leiden University Science and Technology Research Center, was used for bibliometric analysis in this study. VOSviewer presents citation networks, journals, countries, co-author relationships, and co-occurrence of keywords used in documents through visual networks [32]. In this study, the bibliometric analysis of the literature on machine learning and fire safety considered absolute citation counts and the normalized citation count metric, which balances publication year and field differences. The primary reason for choosing VOSviewer is its ability to process normalized citation data, which enables accurate identification of practical and high-quality studies in the literature [32].

3.1. Most Relevant Sources

Academic studies were published in a total of 165 different sources. The number of citations received by these journals is limited to at least 10, and the number of articles published on the subject to at least 3, and the 13 most influential journals are shown in Figure 4. According to the analysis results, the Journal of Building Engineering (16 articles, 257 citations), Fire Technology (12 articles, 210 citations), and Sensors (10 articles, 213 citations) are the most influential journals in this research area. In addition, the fact that three articles published in the journal Sustainability received 112 citations gives an idea of the quality of the articles and their contribution to the field of research. These 13 journals cover interdisciplinary research in fire engineering and safety, civil engineering, artificial intelligence and computer science, electronics and sensor technologies, environmental science, and sustainability.

3.2. Most Frequent Author Keywords

Keywords in studies are critical for rapidly understanding a research topic and its content. As a result of the limitation of using three or above for the analysis of high-frequency keywords in the research field, 45 keywords were determined and visualized with the Sunburst Chart (Figure 5). According to the chart, deep learning (58) is the most used author keyword, while machine learning (34) and fire detection (33) are the second most used words.

The network map was arranged to determine the co-use of keywords and their relationships (Figure 6). Different color clusters in the network represent the co-use of keywords and give an idea about the research areas. Words such as “fire safety,” “fire,” “sensors,” “internet of things (IoT),” “artificial neural networks,” and “transfer learning” in the red cluster give the idea that innovative technologies focus on real-time fire detection, evacuation modeling and intelligent building automation applications in fire safety and firefighting. The keywords in the green cluster, “internet of things (IoT),” “artificial neural networks,” “smart firefighting,” “fire evacuation,” and “computational fluid dynamics (CFD),” represent multidisciplinary research areas such as AI-based evacuation and detection systems, IoT applications, and simulation-based modeling. These studies specifically investigate the development of fire safety technologies and the generation of predictions for building safety through machine learning.

Keywords in the blue cluster, such as “machine learning,” “fire detection,” “data fusion,” “multisensor,” and “time series classification,” indicate that work is being performed in areas such as machine learning-based systems, sensor technologies, and data fusion for early fire detection and prevention. These studies support technological innovations aimed at detecting fires earlier and more reliably. The yellow cluster words “buildings,” “emergency evacuation,” “lightweight network,” “object detection,” and “virtual reality” indicate approaches that utilize innovative network systems, virtual reality, and object detection technologies to manage emergency scenarios in buildings.

In the purple cluster, “remote sensing,” “neural networks,” “image processing,” “feature extraction,” “convolutional neural networks,” “deep learning,” and “smoky environment” represent the combination of remote sensing, image processing, artificial intelligence, and deep learning technologies. These studies cover applications such as detecting, monitoring, and predicting fires and environmental threats.

3.3. Most Relevant Authors

Author analysis is an important method that evaluates which authors have published the most, received the most citations, and are prominent in terms of a specific impact measure (e.g., normalized citations) in a research field [34,35]. Authors who made significant contributions to the literature, are frequently cited, and shape academic studies with their influential publications, are determined as the main actors in the relevant field by taking into account their scientific productivity, the structure of collaboration networks, and social network metrics [36]. This method guides future research and contributes to understanding and directing research trends.

Normalized citation count is a metric that standardizes the number of citations a publication receives by comparing it with the general citation averages in the field. This metric prevents citation counts from being affected by variations in citation density across disciplines in the field. For example, publications in some disciplines generally receive more citations, while others receive fewer citations. Normalized citation count is used to measure the relative impact of a publication by taking these differences into account [37,38]. In order to measure the degree of correspondence between the theoretical knowledge and the analysis results, a multiple regression analysis was performed with the normalized citation count as the dependent variable and the number of publications, total citation count, and average publication year as independent variables. The regression analysis aims to validate the validity and applicability of the metrics used in this study by revealing the weight of the variables affecting the normalized citation values.

The number of publications, number of citations, normalized citation count, and average year of publications of the 42 most influential authors (minimum number of articles 2, number of citations 15) out of 935 authors with publications in the research field are given in

Table 3. Numerical values of authors contributing to the research.

Authors	Article	Citation	Normalized Citation	Average Year
Wu, X	4	130	9.72	2021.75
Wang, Z	3	83	8.2	2022.67
Abdusalomov, A	2	119	5.81	2020.5
Hu, Z	2	72	4.77	2020
Li, A	2	55	4.54	2022
Yeoh, G H	3	55	4.54	2017.67
Yuen, A C Y	2	55	4.54	2022
Liu, G	2	18	4.42	2023.5
Qu, G	2	18	4.42	2023.5
Ren, L	2	18	4.42	2023.5
Wang, L	2	18	4.42	2023.5
Li, Y	3	49	4.05	2022
Liu, Y	3	49	4.05	2022
Zhang, W	3	49	4.05	2022
Xiao, F	2	32	3.99	2023
Zeng, Y	3	37	3.73	2022.67
Huang, D	2	46	3.12	2019
Shibasaki, R	2	46	3.12	2019
Song, X	2	46	3.12	2019
Kumar, A	2	56	3.09	2020.5
Singh, A	2	56	3.09	2020.5
Su, L	2	49	3.03	2021.5
Zhang, X	2	49	3.03	2021.5
Buffington, T	3	41	2.41	2021
Liu, J	2	20	2.26	2022.5
Chokwitthaya, C	2	41	2.1	2019.5
Mukhopadhyay, S	2	41	2.1	2019.5
Zhu, Y	2	41	2.1	2019.5
Hu, Y	2	20	2.08	2022.5
Li, J	2	20	2.08	2022.5
Ezekoye, O A	2	34	2	2021
Gottuk, D	2	46	2	2022
Hammond, M	2	46	2	2022
Hart, S	2	46	2	2022
Rose-pehrsson, S	2	46	2	2022
Wong, J	2	46	2	2022
Wright, M	2	46	2	2022
Merci, B	2	18	1.8	2020.5
Verstockt, S	2	18	1.8	2020.5
Huang, X	7	160	1.62	2022.71
Bilyaz, S	2	18	1.06	2021
Zhang, T	5	93	1.02	2023

. In order to determine the relationships between the indicators in the table, a multivariate regression analysis is performed, showing the effects of the variables of number of citations, number of articles, and average year of publications on the normalized citation count variable (

Table 4. Regression coefficients of the normalized citation count variable.

Model	Unstandardized Coefficients		Standardized Coefficients	t	Sig. (p)
	B	Std. Error	Beta (ß)
(Constant)	−215.809	66.308		−3.255	0.002
Article	−1.054	0.348	−0.583	−3.026	0.004
Citiation	0.045	0.011	0.789	4.193	0.000
AvYear	0.109	0.033	0.437	3.302	0.002

). Accordingly, the number of articles negatively and significantly explains the normalized citation count variable (ß = −0.583, p < 0.004). The number of citations positively and significantly explains the normalized citation count variable (ß = 0.789, p < 0.0001). The average year of publications positively and significantly explains the normalized citation count variable (ß = 0.437, p < 0.002). As a result, since the variables of the author’s number of articles, number of citations, and the year average of publications indicate the citation potential significantly and highly explain the normalized citation count, the normalized citation values for the qualified studies of the authors give correct results.

The Treemap graph, which lists influential authors based on normalized citation counts, is given in Figure 7. According to the graph, light and large colors represent influential authors. Wu (4 articles, 130 citations), with a normalized citation coefficient of 9.72, stands out as the most influential author in the field of study. Wang (3 publications, 83 citations) and Abdulsalomov (2 publications, 119 citations) are the following.

3.4. Country Collaborations

International collaboration is a critical factor that increases scientific productivity and impact. Bibliometric analyses show that such collaborations are associated with articles with higher citation impact. Studies with international co-authorship have the potential to receive more citations than local studies [39,40]. In addition, international collaboration networks contribute to forming academic trends by increasing research performance [33]. The graph of 16 countries with at least three publications and at least 10 citations out of 49 countries with studies in the field of research is given in Figure 8. Accordingly, China (117 publications, 1236 citations) is the country that contributes the most to the field of study. The United States (30 publications, 398 citations) and South Korea (20 publications, 251 citations) are the countries that lead the field of research. In addition, regarding the quality of the publications, England comes after China (113.49) and South Korea (19.80) with a normalized citation coefficient of 12.22, and Singapore, 11.83. The clusters in different colors in the network map in Figure 9, which shows the academic collaboration network between countries and the intensity of this collaboration, give an idea about the collaboration between countries. Accordingly, China, which contributes the most to the field of research, has the largest network in the field by collaborating with nine different countries. India has conducted studies by collaborating with five countries, the United States with four, Saudi Arabia with four, and Japan with four.

3.5. Most Relevant Studies

In identifying influential publications, 94 were identified as a result of filtering publications with 10 or more citations as a first step. A second filtering was performed among these publications, considering thematic suitability and citation impact. Studies with a normalized citation value of 2.00 or higher were selected in this context. The 40 publications above this threshold value were subjected to content analysis and formed the detailed evaluation group of this study. As a result of the analysis in which the basic categories of each study were classified, it was determined that the influential articles in the research field had a total of six categories. Evacuation and emergency management (13) and fire detection and early warning systems (12) were the categories with the most publications. Fire dynamics and spread estimation (7), fire load and material risk analysis (4), intelligent systems and cyber security (3), and fire prediction and risk assessment (3) constituted the other categories. In addition, brief information was provided about the machine learning subcategories used in the studies and the topics they focused on (

Table 5. Fire safety—machine learning qualified studies.

Category	Machine Learning Subcategory	Scientific Field
Fire Detection and Early Warning Systems	Convolutional Neural Network (CNN) [2,41,42,43,44,45,46]. Fuzzy Logic Reasoning [47]. Neural Network [47]. CNN based on YOLOv3 (You Only Look Once) [48]. Deep Learning [49]. Detection Transformer (DETR) [44]. YOLOX [50]. Probabilistic Neural Network (PNN) [51]. Lightweight CNN [46].	Fire detection by classifying objects and extracting features from image data [41]. Detection of common arc faults in low voltage systems for early detection of electrical fires [47]. Detection of collapsed or burned buildings in difficult disaster areas with UAVs [42]. System that detects fires of different sizes and shapes with high accuracy [48]. A deep learning-based model using dilated convolutions for camera-based fire and smoke detection [43]. A data-driven system that detects construction fires on construction sites [2]. Predictive model for early fire detection in indoor areas, providing faster detection than traditional fire detectors [49]. A faster and more accurate detection mechanism with CNN and DETR model for smoke detection [44]. A system that analyzes the variable structure of smoke and performs precise feature extraction with self-attention and self-cooperation mechanisms to improve smoke detection [50]. Early fire detection system with a CNN based cascade classification model using video surveillance systems and image processing technologies [45]. A low-cost classification model for early fire detection using gas and ambient sensors [51]. A fire detection model suitable for real-time and embedded systems using MobileNetV3 and anchor-free object detection methods [46].
Evacuation and Emergency Management	Recursive Neural Network (RNN) [51]. Convolutional Neural Network (CNN) [46,52] Policy Neural Network [46]. Deep Reinforcement Learning [53]. Cellular Automata Model [54]. Decision Tree [55]. Support Vector Machine (SVM) [55]. Artificial Neural Network (ANN) [55,56,57] Fuzzy Logic [56]. Deep Learning [52,58,59] YOLOv4 Evacuee Detector [58]. Back Propagation Neural Network [60]. Discrete Choice Models [61]. Deep Neural Networks (DNNs) [62]. Q-Learning [62].	Optimization of evacuation times in metro stations with VR video tracking [53]. Optimizing evacuation processes and rapid escape strategies in large-scale public buildings [54]. Model that provides dynamic guidance for emergency evacuation processes and determines the safest escape routes within the building [55]. Analysis of exit width, door design and crowd movement to optimize evacuation processes [56]. Analysis of pedestrian movement patterns to understand individual movement dynamics during evacuation processes and improve evacuation management [58]. Modeling people’s pre-action responses to evacuation in a fire [52]. Estimating evacuation rates to monitor building fire evacuations and analyze human behavior [60]. A lightweight and high-accuracy convolutional neural network model that optimizes evacuation planning by identifying escape signs, doors, and stairs in emergency scenarios [57]. A framework that automatically determines the safest evacuation route based on the current fire situation to optimize evacuation processes [61]. A mobile-supported smart fire evacuation system that determines the safest and shortest evacuation route by taking into account individual characteristics and dynamic environmental conditions [62]. Analyzing individuals’ navigation preferences to understand and model evacuation behaviors and optimizing evacuation strategies with agent-based simulations [59]. A system that detects emergencies, estimates individual locations, predicts fire spread and offers the safest evacuation route to the user via an IoT-supported network [63]. A real-time fire evacuation system that can adapt to dynamic environmental changes using graph-based networks and IoT sensors [64].
Prediction of Fire Dynamics and Spread	Decision Tree [65]. Neural Network [65]. Monte Carlo Simulation [65]. Convolutional Neural Network (CNN) [66,67]. Gated Recurrent Unit (GRU) [65]. Deep Neural Networks (DNNs) [68]. Artificial Neural Network (ANN) [69]. Deep Learning [67]. Transfer Learning (TL) [70]. Long Short-Term Memory (LSTM) [70].	Estimating the risk of collapse of steel-framed buildings under fire [65]. A system that predicts smoke spread inside a building in real time [66]. GRU-based deep learning model to predict fire source and intensity inside a building [68]. A model that predicts fire dynamics faster and more efficiently by using ANN and TL to mimic temperature outputs produced by FDS [69]. Detecting signs of flashover and predicting when a fire will reach its flashover point by monitoring its growth through experiments in a real room [67]. Estimation of temperature changes in fire using distributed optical fiber sensors [70].
Fire Load and Material Risk Analysis	Artificial Neural Network (ANN) [71]. Monte Carlo Simülasyonu [71,72]. Extreme Learning Machine (ELM) [73]. Enhanced Cat Swarm Optimization (ECSO) [73]. Genetic Algorithm (GA) [74].	Building damage assessment by estimating the compressive strength of geopolymer concrete exposed to high temperatures [71]. Thermal management and cooling optimization to reduce thermal runaway and fire risks in lithium-ion battery systems [72]. ELM-ECSO model to predict the compressive strength of geopolymer concrete exposed to high temperatures [73]. A model to predict pyrolysis kinetics to assess fire risks of different insulation materials in buildings [74].
Intelligent Systems and Cyber Security	Deep Learning [75,76]. Convolutional Long-Short Term Memory [10]. Artificial Intelligence Engine [10].	Creating augmented digital twins that include building fire safety components by combining image recognition and laser scanning techniques [75]. Artificial-Intelligence Digital Fire (AID-Fire) system that integrates digital twin, IoT sensor network and deep learning techniques to provide real-time fire detection and analysis [10]. A semantically linked data management system for the optimal positioning of intervention equipment and escape routes at the fire site [76].
Fire Forecast and Risk Assessment	Artificial Neural Networks (ANNs) [77,78,79]. Generative Adversarial Networks (GANs) [79]. Feature Ranking [78].	Introducing a new computational framework that improves performance predictions for modeling human-building interactions in building design processes [77,78]. An integrated framework for predicting fire impacts and minimizing fire risk [79].

). Convolutional neural network (CNN) was the most common method, and it was used nine times. This method has been preferred in image-based fire detection and smoke detection studies. Other popular methods, such as deep learning and artificial neural networks (ANNs), have been used six times each. Deep learning methods have been adopted more in fire prediction and early warning systems. Methods such as Monte Carlo simulation, decision tree, support vector machine (svm), and deep neural networks (DNNs) have a medium usage frequency. These methods have been used in fire spread prediction and risk analysis studies. Advanced AI methods such as transfer learning, generative adversarial networks, and Q-learning have been preferred in studies focusing on unique scenarios. Although YOLO-based object detection algorithms (YOLOv3, YOLOX, YOLOv4) have been preferred in a limited number, they have played an active role in the design of fire detection and evacuation systems.

Bibliometric and content analyses show that machine learning applications in fire safety are concentrated in specific thematic areas. However, it is also noteworthy that the literature in this field is still limited in some critical dimensions. In particular, no machine learning-based studies analyze the effect of architectural design parameters (facade geometry, interior space planning, opening locations, etc.) on fire dynamics. Similarly, material-based fire risk classification, prediction systems that work with real-time data streams, evacuation decision support systems based on user behavior models, and integrating multi-layered data sources (sensors, images, air quality data, etc.) have found minimal space in current research. These methodological gaps point to important research opportunities that could be explored in the future, particularly in terms of interdisciplinary data integration approaches and performance-based fire safety design.

Machine learning models tend to overfit when the training data are limited, unbalanced, or non-representative. This situation causes the model to memorize the training data, lose its generalizability, and produce results with low accuracy on new data. In studies, various preventive strategies have been applied to address the overfitting problem encountered during the training process of machine learning models. Among the primary methods commonly used in image processing and classification-based models, data augmentation, dropout layers, early stopping, and k-fold cross-validation techniques stand out [53,58,60,66,67,68,77,78,79].

CNN and DNN-based fire detection models tend to memorize the training data when training with a limited number of images. In order to mitigate this problem, data augmentation techniques have been applied, especially in image-based models. Within the scope of these methods, operations such as rotation, scaling, flipping, brightness and contrast adjustment, and color space transformations (color jittering) are applied to the training images to train the model with more diverse and representative data. Thus, it is intended that the model not only adapts to limited data samples but also produces successful results in different scenarios. In addition, in some studies, an early stopping strategy has been applied to stop learning when the error on the validation set starts to show a particular threshold value or an increasing trend during the training process [41,42,53,68].

In methods such as SVM, decision trees, and ANNs, hyperparameter optimization, pruning, and regularization were used to limit model complexity. Hyperparameter optimization allows the systematic determination of settings that improve the model’s overall performance, while pruning simplifies the model by eliminating unnecessary model branching. Regularization methods prevent overlearning by limiting certain weights in the model’s learning process. In addition, in some deep learning-based studies, dropout layers have prevented the model from focusing only on specific examples or neural pathways [45,52,62,71,77,78,79]. These techniques have significantly contributed to improving the accuracy of machine learning models and making them more resilient to real-world data.

4. Discussion and Conclusions

This study is one of the first to analyze machine learning-based approaches in the fire safety literature comprehensively. The integrated use of bibliometric analysis and qualitative content analysis revealed the quantitative trends and contextual orientations of publications in the field. Thanks to its capacity to process normalized citation data, the VOSviewer tool was used to highlight high-quality publications that not only have high citation counts but also have a high impact on the field. The matches between the most frequently used machine learning models (e.g., CNN, ANN) and fire safety problems were systematically analyzed, detailing which data types and problem categories these models preferred. Prominent machine learning models were identified as keyword clusters and recurring themes in qualified publications. In addition, the clusters revealed in the keyword network analysis and the subcategories defined by content analysis overlap to a large extent. Such cross-validation shows that the trends and conclusions reached are not artificial but derive from real research practices in the field. In addition, methodological depth is added by including preventive strategies against common modeling problems such as overfitting.

Extensive keywords were used to access literature studies integrating fire safety design with machine learning. The words ‘forest,’ ‘wildfire,’ ‘medical,’ ‘cigarette,’ ‘car,’ and ‘road’ were removed using the ‘not’ function of WoS to eliminate publications outside the scope of the research. Only 250 publications were found, the first of which was published in 2001. Furthermore, only 43 articles have been published up to 2019. This situation shows that the integration of machine learning in fire safety has room for development despite the increase in the number in recent years. In addition, the fact that China has approximately 50% of the studies highlights the lack of international contribution and cooperation in the research field.

According to WoS categories, publications have emerged in six categories. Electrical-Electronics Engineering, Computer Science, and Artificial Intelligence Sciences provide an idea that machine learning applications are shaped in the context of fire safety at the level of technical infrastructure and algorithm development. Secondly, the prominent Civil Engineering and Construction Building Technology sciences show that integration studies into direct field applications such as fire detection, evacuation simulations, and structural risk analyses are concentrated. Engineering Multidisciplinary and Material Science Multidisciplinary sciences show that studies are not limited to a single engineering field and allow for multidisciplinary studies. In addition, the Journal of Building Engineering, Fire Technology, Sensors, and Sustainability journals have provided the most effective publications in terms of article and citation numbers in the field of research. These journals accept multidisciplinary studies in engineering fields.

Considering the low number of citations in disciplines with low publication numbers and the potential of newly published studies to be cited over time, normalized citation metrics were used to evaluate author and article rankings more fairly. A regression analysis was performed to confirm the validity of this hypothesis. As a result of the analysis, the number of articles, citations, and the yearly average variables of publications showing the citation potential explained the normalized citation count significantly and highly. According to the normalized citation count, the most influential authors in the field of research are Wu, Wang, and Abdulsalomov. The articles published by Gu et al. [55], Ren et al. [47], Chen et al. [64], and Frizzi et al. [41] were the most influential publications, respectively. China and the USA are the countries that contribute highly to the field of research in terms of the number of articles and the high normalized citation count. The fact that researchers in China and the USA play central roles in national and international collaborations contributes to the increase in scientific productivity in these countries and the achievement of high-impact values in citation networks. This is one of the important factors that reinforce the decisive position of these countries in the field of research.

The keyword and qualified article analysis revealed the formation of subcategories and research trends. The keywords “fire detection, smoke detection, CNN, image processing, lightweight network, data fusion” in the blue cluster indicate early fire detection designs and overlap with the fire detection and early warning systems subcategory. The purple cluster’s keywords “remote sensing, neural networks, image processing, feature extraction, convolutional neural networks, deep learning, smoky environment” represent the fire detection and early warning systems subcategories. The yellow cluster “Buildings, emergency evacuation, lightweight network, object detection, virtual reality” primarily represents keywords suitable for evacuation and emergency management, intelligent systems, and cybersecurity categories. The keywords “smart firefighting, fire evacuation, CFD, IoT, ANN” in the green cluster include keywords belonging to evacuation and emergency management and fire dynamics, and spread prediction categories. The keywords in the red cluster are “fire safety, fire, sensors, internet of things (IoT), ANN, transfer learning, simulation,” representing the categories of fire prediction and risk assessment, and smart systems and cybersecurity.

As a result of the publications’ content analysis and keyword analysis, it was determined that research in engineering disciplines is spread across different themes and supported by interdisciplinary studies. The findings clustered the studies around six main themes: evacuation and emergency management, fire detection and early warning systems, fire dynamics and propagation prediction, fire load and material risk analysis, intelligent systems and cyber security, and fire prediction and risk assessment. The most popular categories were evacuation, emergency management, fire detection, and early warning systems. This reveals that academic interest in human behavior analysis and early fire detection in fire safety has increased in recent years. The studies in six categories are generally divided into methods and results within the framework of different prediction possibilities provided by ML subcategories such as CNN, ANN, SVM, DNN, DNN, YOLO, and DT. CNN is the most widely used ML category. The high accuracy of CNNs in image processing-based fire and smoke detection scenarios explains this preference rate. Deep learning and ANN models came in second place. These methods are preferred in complex data processing applications, such as fire prediction and early warning systems.

Different subcategories of machine learning-based predictions have been used to align with the aims and objectives of the studies. CNNs classify by extracting features such as edges, textures, and shapes in images. Thus, they enable early fire prediction by automatically recognizing visual cues like smoke and flame. CNN requires large data sets, and the training process is time-consuming [53,57]. ANNs process input data through layers, learn patterns in the data, and predict the result. They estimate fire risk by obtaining data such as temperature, smoke density, and gas levels in the building through sensors [77,78,79]. Deep learning learns directly from raw data using multi-layered artificial neural networks. It works incredibly effectively with large, complex data sets such as images, audio, text, and sensor data. It offers powerful and flexible solutions for detection, prediction, and damage classification in fire safety. However, it is prone to overfitting without adjustment and may underperform against new data [74,76]. Decision tree concludes by dividing the data into branches like a tree structure according to its features. It divides the data according to the node features in the structure, and the leaves show the predicted class. It provides the opportunity to simplify critical decision mechanisms by guiding according to sensor data such as temperature, humidity, smoke level, fire prediction, crowd density, exit distance, etc. SVM separates data into different classes with the widest margin by finding the boundary line (hyperplane) separating the data. DT can learn data in too much detail (overfitting); without pruning, it can underperform against new data. It is used in cases such as the detection of anomalies and the distinction between real fire and false alarms in fire detection and prediction systems by providing the maximum separation between classes [58]. However, with high-dimensional or noisy data, performance can suffer, and optimization of the model can be complex. DNN is an artificial neural network that can learn highly abstract and complex features in data. It provides early fire detection by making predictions in multidimensional and complex data such as images, audio, and natural language processing. However, the high computational power and lengthy training process can disadvantage projects with limited resources [55]. YOLOv3, YOLOv4, and YOLOX are advanced deep-learning models used for real-time object detection based on the “You Only Look Once” (YOLO) architecture. It detects objects in images quickly and with high accuracy, regardless of their resolution, environment, and thermal degrees. It shows high performance in tasks such as fire detection, smoke detection, and risky object tracking [50,60]. The fact that YOLO can only work with image-based data and the model requires fine-tuning limits its use. To better understand the relative strengths and limitations of each model used in fire safety predictions, a table summarizing these approaches’ main advantages and disadvantages is presented (

Table 6. Pros and cons of ML models for fire safety prediction.

ML Model	Pros	Cons
(CNN)	High accuracy in image-based fire and smoke detection	Requires large datasets and long training times
(ANN)	Adaptable to various input types; widely used in evacuation prediction	Prone to overfitting without proper regularization
(DNN)	Can capture complex nonlinear fire dynamics	Computationally intensive; requires GPU acceleration
(DT)	Interpretable structure; quick decision-making	Overfitting likely without pruning; may not generalize well
(SVM)	Effective in small-sample fire classification problems	Performance decreases with noisy or high-dimensional data
(YOLO)	Real-time detection in visual surveillance systems	Limited to image-based fire applications; needs fine-tuning

).

The integration of research themes and ML categories into fire safety offers various benefits not only for academic knowledge production but also for practitioners. For example, prioritizing evacuation and emergency management and fire detection systems with AI-based approaches enables the development of decision support systems for fire safety engineers. The effectiveness of models such as CNN and DNN from image-based detection systems offers opportunities for designers to integrate early warning systems into the building envelope. Furthermore, the increasing use of AI in fire risk assessment, material safety analysis, and smoke propagation prediction requires policymakers to consider new assessment tools in developing performance-based codes and regulations. The findings of this study offer practical implications at both engineering and policy levels, encouraging interdisciplinary collaboration.

In this study, the contributions of machine learning algorithms to fire safety have been analyzed from a classificatory approach and within the framework of a critical evaluation. How different algorithms (CNN, YOLO, ANN, SVM, DNN, etc.) are applied to specific fire safety problems has been explained. Then, the advantages and limitations of these methods were presented comparatively. Technical challenges encountered in modeling processes (e.g., overfitting) and proposed solutions to these challenges (e.g., data augmentation, dropout, early stopping, hyperparameter optimization) are detailed. Additionally, bibliometric analysis has provided a theoretical foundation for future studies by identifying methodological gaps (e.g., design parameters, material classification, behavioral data, real-time prediction).

Although the concrete reality of machine learning is increasing in usage intensity due to its high potential in built environments, its integration into fire safety still needs to develop, and the number of qualified publications needs to increase. However, this study has found that the future potential of the research area is promising. In this context, researchers can contribute to fire-safe built environments by using machine learning opportunities to predict fire safety risks, especially in the early design phase. Future studies can provide risk predictions and improvements in the design phase by defining, classifying, and presenting all possible fire risks due to building design and user effects, as well as fire extinguishing and evacuation planning. Thus, the risks of fire occurrence or growth during construction and usage can be minimized.