Application of Ground-Motion Model Based on Deep Neural Network and Earthquake Horizontal-to-Vertical Spectral Ratio to the 2024 Noto Peninsula Earthquake in Japan

ABSTRACT This study investigates the applicability of a ground-motion model (GMM) based on a deep neural network (DNN) and earthquake horizontal-to-vertical spectral ratios (EHVRs), as proposed by Pan and Miura (2025), to the ground-motion intensities recorded during the 2024 Noto Peninsula earthquake (Mw 7.5) in Japan. The DNN model was updated by retraining it on the latest strong ground motion flat file from the National Research Institute for Earth Science and Resilience, integrated with a site-specific EHVR database for Kyoshin network and Kiban–Kyoshin network stations. To address the scarcity of large-amplitude records in the training set, a data augmentation technique was implemented by adding artificial random noise to the observed response spectra and EHVR, thereby enhancing the model’s representation of large-amplitude ground motions. The model’s predictive performance was evaluated by comparing observed and estimated response spectra and conducting a detailed residual analysis. The stability and variability of the GMM were further assessed using both the original training data and independent test data from various seismic events in 2024. The results demonstrate that the DNN model trained on the augmented database outperformed the model trained on the original dataset, particularly in near-field regions. Furthermore, the proposed model exhibited high accuracy when applied to test data not included in the training samples, with estimated spectra showing good agreement with observations. These findings confirm the model’s robust generalization capability and its effectiveness in capturing complex site-specific spectral characteristics, including peak shifts, under strong ground shaking.