Nuclear reactor design faces a major bottleneck: high-fidelity simulations are computationally expensive, slowing innovation for Small Modular Reactors and Generation IV designs. NVIDIA's PhysicsNeMo framework offers a solution via AI surrogate models that predict full spatial fields rather than scalar outputs. Using a fuel pin cell as a worked example, the guide demonstrates training a Fourier Neural Operator (FNO) to jointly predict neutron flux and absorption cross-section fields. This physics-aligned approach achieves an R² of 0.97 versus 0.80 for a baseline gradient boosting regressor, because preserving spatial information captures self-shielding effects that scalar descriptors miss. The workflow covers dataset generation with Latin Hypercube Sampling, data preprocessing with PhysicsNeMo Curator, multi-GPU model training, and deployment as an API for interactive digital twins. Code is available on GitHub.
Table of contents
AI-augmented nuclear reactor simulation and designBuilding an AI surrogate model of a fuel pin cellDataset generation: Efficient samplingModel training with PhysicsNeMoTraining Fourier Neural OperatorsBaseline model resultsIntegrating AI surrogatesGoing furtherGetting startedSort: