Ground Motion Prediction for Japanese Subduction Zone Earthquakes Based on Transformer Model

Accurate prediction of ground motion intensity parameters helps enhance the accuracy of post-earthquake intensity rapid reports and seismic risk analyses. Traditional ground motion prediction equations involve subjective model formulations, and existing machine-learning approaches cannot capture feature relationships on a global scale, resulting in limited predictive accuracy. In this study, we develop a Transformerbased framework that incorporates a Feature Tokenizer technique and a CLS token, enabling the model to perform regression tasks while aggregating critical feature information across the entire dataset. Using <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 5, 0 2 0}$</tex> records of subduction-zone ground motions from the Japanese National Institute for Disaster Prevention database spanning 1997 to March 2025, we trained a Transformerbased subduction-zone ground-motion prediction model. The results show that the proposed method significantly outperforms conventional GMPEs and, compared with deep neural network models, achieves superior performance across various evaluation metrics, demonstrating its advantage in ground motion intensity prediction.