Abstract
We present LazyDINO, a transport map variational inference method for fast,scalable, and efficiently amortized solutions of high-dimensional nonlinearBayesian inverse problems with expensive parameter-to-observable (PtO) maps.Our method consists of an offline phase in which we construct aderivative-informed neural surrogate of the PtO map using joint samples of thePtO map and its Jacobian. During the online phase, when given observationaldata, we seek rapid posterior approximation using surrogate-driven training ofa lazy map [Brennan et al., NeurIPS, (2020)], i.e., a structure-exploitingtransport map with low-dimensional nonlinearity. The trained lazy map thenproduces approximate posterior samples or density evaluations. Our surrogateconstruction is optimized for amortized Bayesian inversion using lazy mapvariational inference. We show that (i) the derivative-based reduced basisarchitecture [O'Leary-Roseberry et al., Comput. Methods Appl. Mech. Eng., 388(2022)] minimizes the upper bound on the expected error in surrogate posteriorapproximation, and (ii) the derivative-informed training formulation[O'Leary-Roseberry et al., J. Comput. Phys., 496 (2024)] minimizes the expectederror due to surrogate-driven transport map optimization. Our numerical resultsdemonstrate that LazyDINO is highly efficient in cost amortization for Bayesianinversion. We observe one to two orders of magnitude reduction of offline costfor accurate posterior approximation, compared to simulation-based amortizedinference via conditional transport and conventional surrogate-driventransport. In particular, LazyDINO outperforms Laplace approximationconsistently using fewer than 1000 offline samples, while other amortizedinference methods struggle and sometimes fail at 16,000 offline samples.