Deep-Learning Emulators of Transient Compartment Fire Simulations for Inverse Problems and Room-Scale Calorimetry

Abstract

This work describes a deep learning methodology for “emulating” temperature outputs produced by the Fire Dynamics Simulator (FDS), a CFD software. An array of artificial neural networks (ANNs) is trained to predict transient temperatures at specified locations for a transient heat release rate (HRR) input. These locations correspond to the locations of thermocouples used in an experimental burn structure. In order to build the training set, A Gaussian process (GP) framework is used to develop a generative model that produces random viable HRR ramps. Although this procedure may require thousands of FDS runs to build a sufficient training set, the application of transfer learning can reduce the required number of runs by nearly an order of magnitude. This refers to the process of initially training an ANN to predict the output of the Consolidated Model of Fire and Smoke Transport (CFAST) and then transferring its knowledge to an ANN that learns to predict FDS outputs. CFAST is a much faster model than FDS, so a large training set can be generated quickly. The final state of the ANN trained to emulate CFAST is used as the initial state of an ANN that learns to emulate FDS. The result is a model that produces FDS temperature predictions with a mean absolute error (MAE) of less than 2°C and runs over five orders of magnitude faster than FDS. The emulators are also capable of learning inverse mappings; i.e. for a given temperature output, they can predict the HRR ramp that would cause FDS to produce the temperature response. This ability to invert for the HRR profile is exercised on data collected from eight fire experiments with peak HRRs up to 200 kW, including four propane burner fires, two methanol pool fires, and two n-Hexane pool fires. The model inverts for the experimental HRR with a MAE of 5.8 kW-15.4 kW (11.3%–16.7%) for the burner tests and 5.0 kW–25.5 kW (12.1%–28.6%) for the pool fire tests, with a tendency to underestimate the HRR of the pool fires. Finally, the computational speed of the emulators allows for the incorporation of CFD physics in Bayesian parameter inversion. As an example, this is demonstrated to infer the radiative fraction from experimental and synthetic data in conjunction with reported uncertainties from the FDS Validation Guide.