DEVELOPMENT OF ANN-BASED DATA-DRIVEN GROUND MOTION MODEL FOR AZERBAIJAN USING TEMPORAL EARTHQUAKE RECORDS OF 2022-2024

Tural Babayev ^1* Gulam Babayev ¹ Sonny Irawan ^2* Emil Bayramov ²

• ¹ Institute of Geology and Geophysics, Ministry of Science and Education, Baku, Azerbaijan
• ² Nazarbayev University, Nur-Sultan, Kazakhstan

This paper evaluates the applicability of the Soviet-era ground motion prediction equation (named as A&K-1979), comparing its predictions against a strong-motion dataset of 500 records. Additionally, an Artificial Neural Network (ANN)-based ground motion model (GMM) is developed for Azerbaijan to improve predictions of horizontal peak ground acceleration (PGA) for seismic hazard analyses. GMMs are essential tools in civil and earthquake engineering and machine learning algorithms are increasingly used to address the limitations of linear empirical models by capturing complex, nonlinear ground motion behaviors. The performance of A&K-1979 is assessed using the most up-to-date dataset (2022-2024), while a novel ANN-based approach is introduced for Azerbaijan. A&K-1979-1 (for PGA ≥ 160 cm/s²) overestimates predictions, whereas A&K-1979-2 (for PGA < 160 cm/s²) underestimates them. The ANN model is trained on the dataset of 500 records to predict the horizontal PGA. The model consists of an input layer (earthquake magnitude and hypocentral distance), followed by three hidden layers (32-32-16 neurons), using the Rectified Linear Unit (ReLU) activation function. The model is validated using a testing dataset of 268 records, with performance evaluated based on bias, standard deviation of residuals (sigma), mean absolute error (MAE), root mean squared error (RMSE) and R². The ANNbased GMM achieved a bias of -0.0076, sigma of 0.5971, MAE of 0.4416, RMSE of 0.5972 and an R² of 0.4601. Despite some limitations, the results demonstrate that the ANN-based GMM provides a more reliable alternative to A&K-1979, particularly for preliminary seismic hazard assessments in Azerbaijan.