Hybrid LSTM model with efficient hyperparameter tuning for earthquake magnitude prediction in Turkey

Earthquakes are complex geophysical phenomena that occur deep in the Earth's crust. The sudden release of accumulated energy is instantaneous, making it difficult to predict in advance. Moreover, there are often no obvious leading indicators that can determine precisely when earthquakes will occur. Therefore, this paper investigated using long short-term memory (LSTM) models from deep learning methods to forecast future magnitudes and timing of a potential large earthquake. To improve the performance of LSTM models, we proposed estimating the hyperparameters of LSTM using optimization techniques such as grid search, random search, genetic algorithm (GA), and particle swarm optimization (PSO). The proposed LSTM-Random, LSTM-Grid, LSTM-PSO, and LSTM-GA methods are evaluated using earthquake datasets covering 1990–2023 (March) from seismic activities in Turkiye's first and second-degree earthquake zones. According to the test results, when the hyperparameters of LSTM were optimized with PSO, the resulting model performed the best and showed the smallest error in earthquake magnitude estimation compared to other methods.