Three-Field Neural Operator Surrogates for Coupled Fast-Reactor Multi-Physics Simulation: data, code, and trained models

Datasets, source code, trained model weights, and figures supporting the manuscript "Three-Field Neural Operator Surrogates for Coupled Fast-Reactor Multi-Physics Simulation", submitted to Computer Methods in Applied Mechanics and Engineering (2026). Contents: data/lfr/ — 5,000-sample lead-cooled fast-reactor (LFR) dataset (packed HDF5, 13 MB), generated by a reduced-order Python neutronics + analytic diffusion + scikit-fem structural stack. data/sfr/ — 500-sample sodium-cooled fast-reactor (SFR) dataset (~22 MB), generated by the fully coupled high-fidelity stack: OpenMC 0.15 (neutronics) + OpenFOAM 2512 (CHT) + Code_Aster 15.6 (3-D thermo-elasto-plastic with HT-9 Norton creep and swelling). code/ — PyTorch implementation of JointDeepONet, SeparateDeepONet, MIONet, and Fourier Neural Operator (FNO), plus 19 experiment driver scripts. models/ — Trained weights (~250 MB) for every experiment in the paper (architecture comparison, data-efficiency sweep, coupling ablation, loss-weighting, geometry extrapolation, SFR baselines, LFR→SFR transfer). figures/ — Final figures used in the manuscript (PDF + PNG, ≥300 dpi). Each sample provides three coupled physical fields on a 16 × 20 (r, z) grid: temperature T [K], fission rate φ [1/m³/s], and hoop stress σ_θ [Pa]. See README.md for full layout, dataset schema, and reproduction instructions. Reproducibility note: The compiled manuscript PDF and LaTeX sources are distributed through the journal once accepted and are intentionally not included in this archive.