Bayesian Data Assimilation in Cold Flow Experiments on an Industrial Thermoacoustic Rig

Abstract

We assimilate experimental data from nonreacting flow in the SCARLET (SCaled Acoustic Rig for Low Emission Technologies) test rig using physics-based Bayesian inference. We model the complex geometry of the combustor with a qualitatively accurate one-dimensional low-order network model. At the first level of Bayesian inference, we assimilate experimental data to optimize the parameter values by minimizing the negative log posterior probability of the parameters of each model, given the prior assumptions and the data. At the second level of inference, we find the best model by comparing the marginal likelihoods of candidate models. We apply Laplace's method accelerated with first and second order adjoint methods to assimilate data efficiently. The first order adjoint is used for rapid data assimilation and optimization. The first and second order adjoints are used for inverse uncertainty quantification. We propose six candidate models for the burner and select the model with most evidence given the data. This produces an improved physical model of the rig, with known uncertainties.