Cross-Regional Seismic Hazard Forecasting Across Diverse Tectonic Regimes via Frozen Geological Priors

Cross-regional seismic hazard forecasting, deploying a model trained in one tectonic setting to predict hazard in a geologically different region, remains an unsolved problem in operational seismology. Data-sparse regions represent the greatest unmet need for reliable hazard estimates yet lack sufficient local earthquake catalogs for region-specific model training. This work presents three contributions that together address the dataset, architectural, and evaluation gaps that have prevented systematic progress on this problem. GeoSeisML-12 is the first multi-regime ML-ready seismic hazard benchmark, comprising twelve 300×300 km patches across eight tectonic regimes on four continents, integrating magnitude-complete earthquake catalogs, seven static geological feature layers, pre-computed temporal seismicity features, and binary hazard targets across twenty spatial and magnitude threshold combinations. The frozen geological prior GNN explicitly disentangles static geological structure from dynamic seismicity patterns, enabling parameter-efficient cross-regional transfer through adaptation of only 3,141 parameters, 1.6% of the full model, without modifying the pre-trained backbone. The Monte Carlo Patch Cycling (MCPC) framework replaces the single hand-selected transfer experiments that characterise prior work with randomised multi-cycle evaluation across all eligible patch combinations, producing statistically robust transfer degradation estimates with leakage-free prior construction and base rate calibration. Evaluated across 1,299 MCPC cycles spanning five independently trained checkpoints and eight tectonic regimes, the frozen-prior architecture achieves mean transfer degradation of 0.20% relative to in-distribution performance, with maximum degradation of 1.0% across all evaluated checkpoints. Both frozen-prior and naive transfer conditions preserve in-distribution predictive capability with negligible performance loss, and in one of five checkpoints transfer AUC substantially exceeds in-distribution performance, consistent with global pre-training providing stronger regularisation than direct fine-tuning on sparse local catalogs. All dataset and code are publicly available.