Abstract
Simulation-based inference (SBI) is a statistical inference approach forestimating latent parameters of a physical system when the likelihood isintractable but simulations are available. In practice, SBI is often hinderedby model misspecification--the mismatch between simulated and real-worldobservations caused by inherent modeling simplifications. RoPE, a recent SBIapproach, addresses this challenge through a two-stage domain transfer processthat combines semi-supervised calibration with optimal transport (OT)-baseddistribution alignment. However, RoPE operates in a fully transductive setting,requiring access to a batch of test samples at inference time, which limitsscalability and generalization. We propose here a fully inductive and amortizedSBI framework that integrates calibration and distributional alignment into asingle, end-to-end trainable model. Our method leverages mini-batch OT with aclosed-form coupling to align real and simulated observations that correspondto the same latent parameters, using both paired calibration data and unpairedsamples. A conditional normalizing flow is then trained to approximate theOT-induced posterior, enabling efficient inference without simulation access attest time. Across a range of synthetic and real-world benchmarks--includingcomplex medical biomarker estimation--our approach matches or surpasses theperformance of RoPE, as well as other standard SBI and non-SBI estimators,while offering improved scalability and applicability in challenging,misspecified environments.