Deep Learning Model for Predicting Earthquake Magnitude Using Bidirectional LSTM with Attention Mechanism on Historical Seismic Data from 2017 to 2025

This study introduces a deep learning model for predicting earthquake magnitudes using historical seismic data collected from seismically active areas including the Mediterranean basin, Southern Europe, North Africa, and Western Asia spanning the years 2017 to 2025. We developed a bidirectional LSTM network enhanced with an attention mechanism to process multivariate time series data comprising magnitude, depth, latitude, longitude, and encoded regional information. The model focuses on significant seismic events with magnitudes above 4.9, which are particularly relevant for populated regions in these areas. Trained on a dataset of 2000 earthquake records from these tectonically active zones, the model achieved a coefficient of determination (R²) of 0.9837 and an accuracy of 90.82% within a ±0.3 magnitude tolerance. Its prediction for the forthcoming earthquake event indicated a magnitude of 7.01, suggesting a potentially impactful seismic occurrence. These findings highlight the promise of deep learning techniques in seismic forecasting within the Mediterranean-European-African-Asian seismic belt and their potential to enhance early warning systems for vulnerable populations.