A machine learning estimator trained on synthetic data for real-time earthquake ground-shaking predictions in Southern California

Rapid and accurate estimation of ground shaking after large earthquakes is crucial for impact assessment. Traditional empirical Ground Motion Models (GMMs) used for real-time estimations can limit accuracy. This study presents Machine Learning (ML) strategies trained on physics-based simulations from the CyberShake database (Southern California Earthquake Center) to improve ground shaking map estimations. The ML-based estimator (MLESmap) using Random Forest and Deep Neural Networks outperforms empirical GMMs for events compatible with the training data, showing significant error reductions for both synthetic and real historical earthquakes in Southern California.

Marisol Monterrubio-Velasco, Scott Callaghan, David Modesto, Jose Carlos Carrasco, Rosa M. Badia, Pablo Pallares, Fernando Vázquez-Novoa, Enrique S. Quintana-Ortí, Marta Pienkowska, Josep de la Puente