Machine Learning-Assisted Calibration of Fission Gas Parameters in the DART Code for U-Mo Fuel

Abstract

Accurate prediction of fission gas swelling in uranium-molybdenum (U-Mo) fuel is essential for fuel qualification. The primary objective of this work is to calibrate fission-gas-related DART parameters such that the model predictions match experimentally observed fission gas bubble concentrations measured from post-irradiation examination. This work presents a systematic methodology combining machine learning surrogate models with global optimization to calibrate fission gas parameters in the DART fuel performance code. High-throughput DART simulations using Latin Hypercube Sampling explored the DART parameter space for RERTR-5 fuel plates. A neural network surrogate (64-128-64 architecture) achieved exceptional accuracy (RMSE = 0.111, R2 = 0.999), representing 92% error reduction compared to the best tree-based method. Global optimization methods (Evolutionary Algorithm, Efficient Global Optimization) achieved 27% RMSE improvement over previous calibration, substantially outperforming local Pattern Search (18%). Validation against independent fuel plates confirmed parameter transferability with approximately 9% total RMSE reduction across varying fission densities, temperatures, and grain sizes. Bayesian calibration provided uncertainty quantification, revealing that nucleation and re-solution parameters exhibit different identifiability levels, informing priorities for future experiments. This methodology through integrating surrogate modeling, global optimization, and Bayesian uncertainty quantification provides an efficient, reproducible framework for multivariate parameter calibration in mechanistic fuel performance codes.