Model Considerations for Fire Scene Reconstruction Using a Bayesian Framework

Towards the development of a more rigorous approach for coupling collected fire scene data to computational tools, a Bayesian computational strategy is presented in this work. The Bayesian inversion technique is exercised on synthetic, time-integrated data to invert for the location, size, and time-to-peak of an unknown fire using two well-known forward models; Consolidated Model of Fire and Smoke Transport (CFAST) and Fire Dynamics Simulator (FDS). A Gaussian process surrogate model was fit to coarse FDS simulations to facilitate Markov Chain Monte Carlo sampling. The inversion framework was able to predict the total energy release by all fire cases except for one CFAST forward model, a 1000 kW steady fire. It was found that insufficient information was available in the time-integrated data to distinguish the temporal variations in peak times. FDS performed better than CFAST in predicting the maximum energy release rate with the posterior mean of the best configurations being 0.05% and 2.77% of the true values respectively. Both models performed equally well on locating the fire in a compartment.