GeoSeisML-12

GeoSeisML-12 is a multi-regime machine learning benchmark for cross-regional seismic hazard forecasting. It comprises twelve 300×300 km patches spanning eight tectonic regimes across four continents, covering January 2000 to December 2024. Each patch includes: magnitude-complete earthquake catalogs sourced from USGS ComCat with patch-specific Mc estimation, Gardner-Knopoff declustering, and Scordilis (2006) Mw homogenisation; seven static geological feature layers (Vs30, crustal thickness, sediment thickness, elevation, slope, roughness, heat flow, tectonic stress, active fault proximity) rasterised to a common 0.1° grid; seventeen temporal seismicity features computed at monthly timesteps; binary hazard targets across twenty spatial radius and magnitude threshold combinations; and heterogeneous graph objects with embedded chronological train/validation/test splits stored as PyTorch Geometric Data objects. The dataset is designed for spatial seismic hazard forecasting and cross-regional transfer learning evaluation. A frozen GMM-PCA geological prior (k=6 clusters) is included for leakage-free transfer experiments. Primary target: Mw ≥ 3.0 within 50 km within 30 days. Code is available at: https://github.com/rahulr2003/Cross-Regional-Seismic-Hazard-Forecasting