Integrated Seismicity Analysis and Machine Learning Framework for Predicting Earthquake Parameters in the Western United States

Abstract

This study presents a machine learning framework for predicting key seismic attributes—epicentral Latitude, Longitude, and hypocentral Depth—for the Western North American plate boundary. The analysis uses a homogenized United States Geological Survey (USGS) dataset of moderate-to-large earthquakes (M ≥ 5.5) recorded between 1980 and 2024. The methodology incorporates median-based imputation, removal of non-tectonic events, and three feature-scaling strategies: Z-score standardization, Min–Max normalization, and a hybrid standardization–normalization scheme. Three regression families are evaluated across all scaling variants and predictive targets: Ordinary Least Squares Linear Regression, Random Forest Regression, and a Multilayer Perceptron neural network, resulting in 27 experimental configurations.
Results show that Z-score standardization provides the most consistent performance across models. Random Forest Regression achieves the highest accuracy for all targets, with a peak R2 of 0.85 for Depth prediction, substantially outperforming both Linear Regression and neural networks. These outcomes highlight the non-linear nature of seismic attributes and demonstrate the effectiveness of ensemble tree methods, even with limited feature sets.
The findings illustrate that seismic spatial coordinates and depth can be estimated reliably using readily available event metadata, without waveform processing or dense observation networks. The proposed framework is computationally
efficient and suitable for rapid regional assessments, offering a practical bridge between traditional seismic analysis and modern machine learning approaches.