Moment Magnitude Estimation Using Machine Learning Algorithms with an Application to the Western United States

Abstract The moment magnitude (MW) is crucial for risk assessment, seismic hazard analysis, and other seismological and geotechnical studies. In this study, we use various machine learning (ML) algorithms to estimate MW of earthquakes using peak ground acceleration (PGA) 5%-damped pseudo-spectral acceleration (PSA) at 21 different periods, hypocentral distance (Rhypo), the depth to the top of the rupture area (Ztor), fault mechanism (FM), and timed-average shear-wave velocity of the upper 30 m of soil (VS30) as input parameters. In this study, we train and test different ML algorithms, including Support Vector Regression, Kernel Ridge Regression, Random Forest Regression, Gradient Boosting, and Neural Networks. Then, we use an ensemble model to combine the models mentioned. We utilize the NGA-West2 database. We focus on data characterized by MW for both training and testing purposes. Then, we estimate the MW for records where the magnitude type is not specified as MW. Results indicate that by using ML, we can quickly and accurately estimate MW for both new seismic events and the updated existing NGA-West2 flatfile for events that MW is inferred from other magnitude metrics. Additionally, this method can be used to address discrepancies in MW values across different earthquake catalogs.