Efficient posterior inference & generalization in physics-based Bayesian inference with conditional GANs

Abstract

In this work, we propose a conditional generative adversarial network (cGAN) to sample from the posterior of physics-based Bayesian inference problems. We utilize a U-Net architecture for the generator and inject the latent variable using conditional instance normalization. We solve the inverse heat conduction problem and demonstrate how the proposed strategy effectively quantifies the uncertainty in the inferred field. We also show that the structure of the generator promotes generalizability due to the local-nature of the learned inverse map.

Tasks

Reproductions