Rapid basin-scale ground motion mapping with a process-guided neural operator

Data, code, and model checkpoints accompanying the paper. Directory Structure ├── GMNO_main.ipynb # Main analysis notebook├── utils.py # Utility functions├── train.py # Training routines├── models.py # Model definitions (GMNO, MLP, ResNet, etc.)│├── requirements.txt # pip dependencies├── environment.yml # Conda environment definition│├── checkpoints.zip # Trained model weights & RBF-ROM checkpoints│├── Chino_Data/ # Observation data for Case study: 2008 Chino Hills Mw 5.4 earthquake│├── media_30.pkl # Velocity model (Vp, Vs, density) processed for 30×30 grid├── resampled_dem_tensor_60.pt # DEM at 60×60├── resampled_dem_tensor_120.pt # DEM at 120×120├── resampled_dem_tensor_30.pt # DEM at 30×30 Notebook **`GMNO_main.ipynb`** is the main entry point. It covers: 1. **Data loading & preprocessing** — 3D seismic simulation database (5000 scenarios, 30×30 km)2. **Model training** — GMNO (proposed) + baselines (RBF-ROM, MLP, ResNet)3. **Batch evaluation** — R², MAE, MAPE, SSIM vs. source parameters and fault types4. **Data efficiency analysis** — Metrics vs. training set size5. **Inference efficiency** — Speed-accuracy trade-off with bubble charts6. **Generalization tests** — Out-of-distribution source depth / orientation7. **Case study — 2008 Chino Hills** — Comparison with BSSA14 GMPE, ShakeMap, and Kriging8. **Assimilation** — Fine-tuning with observed ground-motion data9. **Land cover analysis** — Performance across different land-use types10. **Uncertainty quantification** — Source parameter perturbation