Bayesian Inference in Physics-Based Nonlinear Flame Models

Maximilian L. Croci, Ushnish Sengupta, Matthew P Juniper

Keywords: Bayesian inference, deep learning, neural networks, inverse problems

TL;DR: Bayesian neural network ensembles are used to infer the parameters and uncertainties of a nonlinear flame model from experimental images.

Abstract: This study uses a Bayesian machine learning method to infer the parameters of a physics-based model of a bluff-body-stabilised flame in real-time. An ensemble of neural networks is trained on a library of simulated flame fronts with known parameters, generated using a level-set solver, LSGEN2D. The ensemble learns a surrogate of the approximate Bayesian posterior of the parameters given the observations, from which the flame can be re-simulated beyond the observation window of the experiment. The method is general: once trained, the ensemble can be used to infer the parameters from any bluff-body-stabilised flame as long as the flame is qualitatively similar and the parameters lie within the ranges in the training library. Amortized inference takes milliseconds, which is fast enough to work in real-time.