<b>Dataset and Source Code for "</b>Physics-Corrected Boundary Neural Operators for Three-Dimensional Fractional Thermoelasticity with Curved Boundary Elements and Hierarchical Solvers<b>"</b>

The free-form three-dimensional benchmark contains 30 geometries split into 18 training, 4 validation and 8 test geometries, with all internal-cavity geometries withheld from training.</p><p dir="ltr">Representative results show that eight boundary-integral correction iterations reduced the mean FNO boundary-flux error from 6.513% to 0.589% on unseen two-dimensional geometries. In the free-form three-dimensional test set, four GMRES iterations reduced the mean flux error to 0.3757%, while a 24-step causal rollout achieved an aggregate flux error of 0.1908%.

At N=12,000N=12{,}000, compressed Kelvin displacement and traction operators achieved storage reductions of 28.06× and 25.20×, respectively, with sampled matrix–vector multiplication errors below 0.2%.</p><p dir="ltr">The repository is intended to support reproducibility and reuse of the computational methodology. It includes or is designed to include benchmark definitions, geometry and split information, numerical data underlying figures and tables, trained-model checkpoints, sum-of-exponentials configurations, boundary-integral and Kelvin–Somigliana solver components, hierarchical matrix routines, configuration files, scripts for reproducing reported results, and environment information.

The manuscript treats the demonstrated capabilities as complementary modules rather than claiming a single fully integrated large-NN, topology-diverse fractional thermoelastic solver.</p>