F-FNO basin-scale tsunami surrogate model (code, weights and training data sample)

Training, inference, and evaluation code for a Factorized Fourier Neural Operator (F-FNO) surrogate model of basin-scale tsunami propagation.

Companion archive for: Kim et al., "A Factorized Fourier Neural Operator Surrogate for Basin-Scale Tsunami Propagation", Geoscientific Model Development, 2026.

Version 2.0.0 accompanies the revised manuscript. The scenario database was regenerated with a stratified Latin hypercube design, the model was retrained on it, and the code and data of the additional experiments of the revision were added. 

This archive contains two files:

1. ffno-tsunami-v2.0.0.zip (~1.87 GB) — Source code, model weights, and scenario tables
  - train.py: full training code (architecture, loss, training loop)
  - inference.py: autoregressive rollout inference and diagnostics
  - convert_comcot_to_nc.py, split_loader.py: COMCOT output conversion and split loading
  - scenarios/: Latin hypercube scenario parameter table (864 scenarios) and train/val/test split list files
  - weights/: the Selected model and the three configuration variants, the four leave-one-source-out folds, the magnitude-interpolation holdout, and the model of the original submission used in the design comparison
  - comcot/: COMCOT control file template and input generation scripts
  - gp_baseline/: Gaussian process baseline code and gauge peak tables
  - bathymetry_perturbation/: perturbed depth files and scripts of the bathymetry experiment
  - grid_refinement/: grid convergence case generation and analysis
  - logs/: training logs of the six runs
  - README.md: folder map, run commands, and package versions

2. test_EM_Ep01_{WC94,ML10}_0NN.nc (54 files, ~41.1 GB in total) — Test-EM evaluation dataset
  - One NetCDF file per scenario for the most challenging test split (unseen epicenter and extrapolated Mw 8.0), drawn from the Latin hypercube database
  - Each file holds the free-surface elevation, the depth-averaged velocities, and the bathymetry of one scenario on the full grid
  - The files are provided individually so that a single scenario can be downloaded without the whole set
  - Sufficient to reproduce all Test-EM results reported in the paper

The full training dataset (~640 GB, 864 scenarios), the fine-grid simulations, and the perturbed-bathymetry reruns can be regenerated from the provided scenario tables, depth files, and scripts using COMCOT v1.7.