REAL-TIME PARAMETER INFERENCE OF NONLINEAR BLUFF-BODY-STABILIZED FLAME MODELS USING BAYESIAN NEURAL NETWORK ENSEMBLES

We assimilate the parameters of a low order physics-based model of a bluff-body-stabilized premixed flame by observing OH PLIF and PIV images of a 1.1 MW flame. The model is a five parameter level set (G-equation) solver with a prescribed velocity field. A Bayesian ensemble of neural networks (BayNNE) is trained on numerical simulations of the model at 2400 different parameter combinations. Once trained, the BayNNE observes the experimental data and outputs the expected values and uncertainties of the parameters of the model that best fits the experimental data. Using this model, we extrapolate the heat release rate field in physical space beyond the observed window in the experiments, and in parameter space to smaller perturbation amplitudes. We then convert the periodic heat release rate field into a distributed n − τ model, which we enter into a thermoacoustic Helmholtz solver. We find that the thermoacoustic eigenvalue drift is small but measurable, is stabilizing, and does not vary significantly during the experimental run or with the velocity amplitude. This is primarily because the time delay field τ, which is determined by the convection speed, is similar for all cases. This is consistent with the experimental images, which exhibit intermittent bouts of thermoacoustic oscillations that die away. Although this paper’s conclusions for thermoacoustic behaviour are unsurprising, the method it describes is a potentially cheap way to combine sparse experimental measurements with copious low order simulations.