Bayesian Optimization for Enhanced Prediction of Earthquakes with Machine Learning Techniques

Abstract

Earthquakes are destructive natural hazards that strike suddenly and unexpectedly, caused by elastic strain energy released along faults, mainly resulting from the gradual accumulation and persistent movement of tectonic plates. Several new approaches have been suggested to predict earthquakes. Machine learning (ML) methodologies have recently emerged as a robust mechanism in dealing with huge, complex, nonlinear, and less dependent on stringent assumptions. Most studies are often too narrow model comparisons, systematic hyperparameter tuning, and remain geographically constrained. Comprehensive benchmarking of regression-based machine learning and Bayesian tuning remains scarce, especially for the high seismicity regions of northern and eastern Iraq. multiple Machine learning methods were employed to model the earthquake magnitude in Iraq for the period 2004-2024. Six Bayesian optimization methods were implemented across the ML methods to optimize the model hyperparameters for the models. Eight mathematical indicators of seismicity parameters are extensively used as input features for the target defined as the earthquake magnitude observed prediction. The Extra Trees Regressor method demonstrated dominant predictive capability among the evaluated metrics, after parameter optimization using the Bayesian Optimization Hyperband. Magnitude deficit, or the difference between the largest observed magnitude and the largest expected magnitude based on the Gutenberg-Richter relationship, had the key influence on earthquake magnitude, as notably by the analysis of variable importance for the eight seismic parameters. These findings provide clear evidence of the effectiveness of Extra Trees Regressor in elucidating the complex relationships underlying earthquake magnitude and provide significant insights to guide the development of data-driven strategies aimed at improving earthquake response in Iraq.