PDE-based Bayesian Inverse Problems in All-at-once Framework

Abstract

We propose a modern all-at-once framework for PDE-based Bayesian inversion by introducing the PDE residual as probabilistic information. In this framework, the PDE residual is treated as a soft constraint with a residual covariance and is directly incorporated into the likelihood function. The new hybrid likelihood function eliminates PDE solves, thus reducing expensive computational costs, while performing rigorous Bayesian inference. Theoretically, we establish the well-posedness of the posterior with uncertainty quantification, and numerically, we propose two efficiency inversion algorithms based on two posterior sampling methods. In particular, one algorithm is based on Metropolis-within-Gibbs MCMC method with a PDE-informed block- diagonal Gaussian prior, and the other is based on a joint sampling strategy. Numerical experiments for a semiconductor inverse problem confirm that the proposed Bayesian all-at-once approach yields a speedup of approximately a factor of two, while having a comparable reconstruction quality as the conventional reduced-based setting.