Brain-STORM: A deep learning model for computationally fast transient high-rise fire simulations

Abstract This paper presents a deep learning model for quickly predicting the temporal evolution of smoke, temperature, and pressure during a high-rise fire scenario. The deep learning model, titled Brain-STORM, serves as a fast and accurate surrogate model of Fire-STORM, a previously developed physics-based fire model. To build the training set, a Gaussian process generative model is used to produce randomly drawn simulation inputs that are designed to resemble potential use cases. These inputs are run in Fire-STORM and the results are used to train Brain-STORM to predict the output that Fire-STORM would produce for new input cases. Once trained, Brain-STORM produces nearly identical results to Fire-STORM and runs over 100 times faster. To evaluate the accuracy of Brain-STORM for realistic use cases, Fire-STORM and Brain-STORM results are compared for multiple cases involving heat release rate curves that are not present in the training set. These curves include a t-squared fire ramp and a fire ramp from NIST's single office workstation fire tests. Finally, given that Brain-STORM can learn inverse mappings directly, the utility of Brain-STORM as a fast heat release rate inversion tool is evaluated.