Bayesian formulations represent one of the prominent approaches for addressing problems of model calibration. Existing Bayesian methodologies are hampered by the high-dimensionality of unknown model parameters and the high computational cost for inference. The present paper advocates a Variational Bayesian inference engine which exploits derivative information available from deterministic adjoint formulations. Furthermore we propose sparsity-enforcing priors that are suited for spatially-varying model parameters and a greedy algorithm for learning the associated basis set.      «

Bayesian formulations represent one of the prominent approaches for addressing problems of model calibration. Existing Bayesian methodologies are hampered by the high-dimensionality of unknown model parameters and the high computational cost for inference. The present paper advocates a Variational Bayesian inference engine which exploits derivative information available from deterministic adjoint formulations. Furthermore we propose sparsity-enforcing priors that are suited for spatially-varying...     »