Surrogate-Guided Generative Modeling with Physics Verification for Expensive Multiphysics Inverse Problems

Abstract

High-fidelity multiphysics simulations provide physically rich information for nuclear reactor analysis, but their computational cost often limits the duration and diversity of conditions that can be practically simulated. This work introduces a physics-informed guided generative modeling framework that learns from limited high-fidelity computational fluid dynamics (CFD) data and expands them into a searchable distribution of candidate states for coupled-physics inverse problems. Tabular generative models learn assembly-level spatial patterns from high-fidelity CFD simulations, while a differentiable surrogate guides candidate generation toward improved responses from a downstream neutronics model. Importantly, the surrogate is used only for sampling-time guidance; every generated candidate is independently evaluated by SIMULATE3 (a neutronics solver), ensuring that optimization decisions remain grounded in verified physics calculations. Two strategies are developed to balance prior fidelity, exploration, and computational cost: a hybrid explorer--generator framework and a more economical rank-weighted generative framework. The methodology is demonstrated for assembly-level DLT field inference for fuel assembly bowing using the BEAVRS benchmark. Starting from a baseline detector root-mean-square (RMS) error mismatch of approximately 5.1%, brute-force evaluation of the CFD-informed dataset alone achieved 4.2%, whereas generative optimization reduced the best verified RMS to 1.86%. The rank-weighted TabDDPM framework reached 2.08% using substantially fewer SIMULATE3 evaluations, while TabDiff reached 3.12\% with greater spatial smoothness. The results further show that improved detector agreement requires controlled exploration beyond direct reproduction of the CFD prior. Overall, the proposed framework establishes a general strategy for combining limited high-fidelity data, generative exploration, surrogate guidance, and continuous physics verification to enable computationally tractable inverse analysis of expensive multiphysics systems.