Conditioning-dependent variational inference for uncertainty quantification in physics-based inverse problems

Summary

This work develops variational-autoencoder-based posterior approximation for Bayesian inverse problems in heat conduction and benchmarks the resulting posteriors against Markov chain Monte Carlo (MCMC) reference solutions. A Fisher-Information-Matrix-based analysis explains how the conditioning of the inverse problem determines when amortized variational posteriors under- or over-estimate uncertainty.

Preprint: arXiv:2609.25145

Recommended citation: Srivastava, A., Hazra, A., & Dubaku, R. Conditioning-dependent variational inference for uncertainty quantification in physics-based inverse problems. Under review, Probabilistic Engineering Mechanics. arXiv:2609.25145.
Download Paper