Adaptive multi-fidelity surrogate modeling for Bayesian inference in inverse problems


主講人:周濤 中國科學院




內容介紹:The polynomial chaos expansion is widely used as a surrogate model in the  Bayesian inference to speed up the Markov chain Monte Carlo calculations.  However, the use of such a surrogate introduces modeling errors that may  severely distort the estimate of the posterior distribution. In this talk, we  present an adaptive procedure to construct a multi-fidelity polynomial  surrogate. More precisely, the new strategy starts with a low-fidelity surrogate  model, and this surrogate will be adaptively corrected using online  high-fidelity data. The key idea is to construct and refine the multi-fidelity  surrogate over a sequence of samples adaptively determined from data so that the  approximation can eventually concentrate to the posterior distribution. We also  introduce a multi-fidelity surrogate based on the deep neural networks to deal  with problems with high dimensional parameters. The performance of the proposed  strategy is illustrated through two nonlinear inverse problems.