Abstract:

Significance Firefighter protective clothing serves as a critical barrier for firefighters operating in extreme thermal hazard environments. Consequently, accurate prediction and optimization of the thermal protective performance of firefighter protective clothing present a core scientific challenge. The physical experiments provide direct performance test data, but their applications are restricted by destructive nature, high cost, long duration, and the difficulty in replicating complex dynamic conditions. Numerical simulations gain computational efficiency but face bottlenecks from their reliance on precise boundary conditions and high computational costs as models grow more complex. The combination of these bottlenecks necessitates a new research paradigm and data-driven machine learning provides promising solution. These algorithms enable learning of high-dimensional mapping relationships from large-scale experimental or simulation data. They could therefore predict outcomes without directly solving complex systems of heat and mass transfer differential equations. This data-driven approach demonstrates high possibility of effectively overcoming the efficiency and accuracy limitations that traditional methods face when dealing with complex dynamic conditions, and shows immense potential for performance assessment of firefighter protective clothing and the dynamic prediction of human burn injury risk.

Progress In machine learning algorithms, the attributes of data instances are referred to as ″features″. The input features cover three major dimensions, which are fire environment parameters, physical properties of fabrics, and human physiological indicators. The output response comprises quantitative measures of the performance of firefighter protective clothing, such as the second-degree skin burn time. Machine learning algorithms could refine the mapping relationship between input features and output responses, breaking through the limitation of linear assumptions. The data sources for machine learning algorithms are derived from physical experiments, literature compilations, and numerical simulations. The precise identification and engineering of key features help to improve the performance of machine learning models. The black-box nature of machine learning algorithms significantly reduces time costs and improves computational efficiency, but input data of poor quality may cause the model to produce biased results. The performance of these algorithms is evaluated based on the accuracy of prediction results, goodness of fit, and stability, with studies demonstrating that machine learning models outperform empirical equations. Furthermore, machine learning facilitates an emerging research direction, i.e. intelligent inverse design, which employs algorithms to find optimal fabric parameters that satisfy specific protective performance requirements. This approach offers a framework for the intelligent inverse design of firefighter protective clothing.

Conclusion and Prospect The prediction and design methods for the thermal protective performance of firefighter protective clothing are undergoing an evolutionary process, transitioning from physical experiments and numerical simulations toward data-driven and intelligent directions. At present, data-driven research has made practical advancements, but several challenges remain. First, collecting data from firefighters' live burn exercises involves high risks and substantial costs. Future research on the intelligent design of firefighter protective clothing can integrate experimentally validated high-fidelity numerical models or computational fluid dynamics heat transfer numerical simulations to generate virtual training data for multiple working conditions. Second, existing design paradigms lack a fully intelligent inverse design capability tailored to specific protection objectives. Future research could develop personalized inverse design platforms constrained by prior knowledge and driven by data. Finally, machine learning algorithms possess high application value in the areas of performance prediction for firefighter protective clothing and firefighter training. Exploring a synergistic pathway that combines ″physics-informed priors, data-driven models, and artificial intelligence agents″, and constructing digital twin systems, is expected to advance the design of firefighter protective clothing in a more precise, practical, and intelligent direction. This will provide key scientific support for development of intelligent firefighter protective clothing and next-generation safety assurance systems.

Key words: firefighter protective clothing, thermal protective performance, machine learning, performance prediction, reverse design