Data and code for Scheduled Momentum Diffusion Improves Hydraulic Shock Fidelity in Strong-Form Shallow-Water Physics-Informed Neural Networks

Abstract

Companion data and code for "Scheduled Momentum Diffusion Improves Hydraulic Shock Fidelity in Strong-Form Shallow-Water Physics-Informed Neural Networks" by Huicong An and Chaojun Ouyang. The package includes 102 formal shallow-water PINN experiments covering nine dam-break depth ratios (2:1 to 10:1), diffusion ablations, bounded Fourier/RAR/weak-form comparisons, a strong learned bed step, radial two-dimensional flow, a matched conservative-residual comparison, alternative diffusion operators, and steady subcritical flow over a smooth bump. Formal comparisons retain all three seeds (2025, 2026 and 2027). Contents comprise source code, resolved physical/training configurations, prediction fields, per-seed metrics, analytical references, three radial FVM reference grids, four Xia2017 smooth-step cross-checks, figure/table source data, and executable reproduction commands. The model archive contains 102 final states and 48 additional states supporting the training-evolution analysis. Configuration and evaluation records for 36 auxiliary calibration/selection runs are retained without their full fields or model weights. The code_and_data archive supports rebuilding all six main and six supplementary figures, generating numerical summaries for all six tables, and recomputing field/reference metrics without retraining. The trained_models archive additionally supports inference checks and hydraulic re-evaluation. PINN and FVM environments are documented separately. Software is licensed under MIT; numerical data, model weights, figures and standalone documentation are licensed under CC BY 4.0. See README.md and the file-level license map for details.