Width-Saturated Fault Scaling and AI-Driven Seismic Hazard: A Global First-Principles Machine Learning Framework

Sujan Bhattarai 

Traditional probabilistic seismic hazard analysis (PSHA) relies on empirical magnitude-area scaling relationships that systematically overestimate energy release in large, geometrically saturated fault systems. This study presents a dynamic, data-driven framework integrating first-principles geophysics with Gaussian Process Regression (GPR) to produce a physics-informed global seismic hazard index across 25 major fault systems. The central empirical result is a seismic moment-fault area scaling exponent of 1.08, significantly below the classical self-similar assumption of 1.5, demonstrating that large fault systems are width-saturated and bounded by the finite seismogenic thickness of the crust. This revised scaling achieves a coefficient of determination (R2) of 0.928 compared to 0.785 for the classical model. The GPR model, trained on K-Nearest Neighbors (KNN)-imputed fault geometry, geodetic coupling fractions, and effective slip rates, independently ranked the San Andreas Fault (Central/Parkfield segment) as the highest-hazard system globally (index = 100.0/100), while correctly assigning near-zero current deficits to recently ruptured faults such as the Japan Trench (Tohoku, 2011 Mw 9.0). No hardcoded thresholds or expert-assigned classes were used at any stage; all ranking emerged from the learned covariance structure of the training data. The framework provides a scalable, threshold-free methodology for infrastructure safety assessment, insurance risk modeling, and national seismic hazard characterization. [287 words]