Combining data-driven and physics-based models for fire applications




Buffington, Tyler Colby

Journal Title

Journal ISSN

Volume Title



Every year, fires inflict billions of dollars of damages and cause thousands of deaths in the United States. As a result, significant resources are allocated to detecting, suppressing, and resisting fires. The costs of fire mitigation efforts contribute to the US fire problem, but ideally these measures have a net positive impact by improving fire outcomes. In many cases, it is difficult to quantify the effect of fire mitigation strategies on fire outcomes. This makes identifying the most effective strategies difficult. This work aims to address this problem through the development of various data-driven and physics-based models. First, it uses historical fire incident data to quantify the relationship between fire service response times and residential fire outcomes in the US. This relationship implies a temporal growth of fires that is mitigated by quick fire service response times, which motivates the development various data-driven models to help fire departments reduce response times. These models forecast spatial and temporal incident frequencies as well as predict the impact of various operational decisions on response times. Although reducing response times is valuable, it represents only one component of the overall effort to mitigate fire outcomes. Other strategies--including fire protection engineering--benefit from physics-based models that allow practitioners to simulate potential fires to evaluate designs. A major limitation of these models is that the most accurate ones are computationally expensive. This work addresses this limitation through the development of deep learning "emulators," which leverage artificial intelligence to produce the output of a transient physics-based model much faster than the model itself. The ability to quickly produce the results from high-fidelity physics-based models is useful for many applications including room-scale calorimetry, which is demonstrated in this work with experimental data. Overall, the models developed in this work aim provide valuable information that helps decision-makers identify the most effective strategies for mitigating the US fire problem.


LCSH Subject Headings