Using machine learning in physics-based simulation of fire

There is a current need to provide rapid, high fidelity predictions of fires to support hazard/risk assessments, use sparse data to understand conditions, and develop mitigation strategies. Machine learning is one approach that has been used to provide rapid predictions based on large amounts of data in business, robotics, and image analysis; however, there have been limited applications to support physics-based or science applications. This paper provides a general overview of machine learning with details on specific techniques being explored for performing low-cost, high fidelity fire predictions. Examples of using both dimensionality reduction (reduced-order models) and deep learning with neural networks are provided. When compared with CFD results, these initial studies show that machine learning can provide full-field predictions 2-3 orders of magnitude faster than CFD simulations. Further work is needed to improve machine learning accuracy and extend these models to more general scenarios.