Evaluating Baseline Machine Learning Models for Earthquake Magnitude Prediction in Türkiye

Özet

Due to its tectonic framework, Turkey faces a constant and severe threat from devastating earthquakes. To mitigate seismic risks, predictive modeling utilizing Artificial Intelligence (AI) methodologies has become increasingly vital. In this study, baseline machine learning models—including Decision Tree, Artificial Neural Networks (ANN), Bayesian Linear Regression, and Linear Regression—were evaluated on the Microsoft Azure Machine Learning Studio platform using historical seismic catalogs (magnitude ≥ 3.5) from the Kandilli Observatory. The performance results demonstrated that models relying solely on historical spatial and temporal parameters yielded insufficient predictive accuracy, with maximum R-squared (R²) scores reaching only 0.249. These baseline findings confirm that classical catalog-based predictions fail to capture the chaotic, non-linear dynamics of seismicity. The study concludes that to achieve reliable forecasting and life-saving early warning mechanisms, future models must transition from simple historical regressions to multidimensional Deep Learning architectures (such as LSTM or XGBoost) that integrate real-time physical constraints, including GNSS strain data and active fault mechanics.