Bayesian inference of deep learning based reduced order model: How does it compare to Bayesian inference of full model?

Abstract

Going from data to build a robust differential/algebraic equation governing the phenomena requires calibration of the coefficients which can have a large range of possible values. The Bayesian inference framework allows narrowing down the range, albeit with a potential snag if the forward model being built is too sophisticated. A reduced order model is needed to make Bayesian inference feasible in such cases, and deep learning in the form of LSTM autoencoders are deployed. We compare the Bayesian inference of a full model to that with a reduced order model to get a proof of concept