Multi-input integrative neural network for soil seismic response modeling at KiK-net downhole array sites

Prediction of the soil seismic response is of primary importance for geotechnical earthquake engineering. Conventional physics-based models such as the finite element method (FEM) often face challenges due to inherent model assumptions and uncertainties of model parameters. Furthermore, these physics-based models require significant computational resources, particularly when simulating seismic responses across numerous soil sites. In this study, a multi- input integrative neural network is developed for predicting soil seismic response based on the recorded data from a large number of downhole array sites. Ground motions, seismic event information, and wave velocity structures of the sites are utilized as input data in the proposed neural network, enabling the model to adapt to various site conditions. Comparative assessments against state-of-the-art FEM models demonstrate that the proposed models exhibit superior prediction performance with increased efficiency. Furthermore, the pre-training technique, a transfer learning method, is employed to predict the seismic response at new stations. By fine-tuning the pre-trained model derived from the extensive dataset with limited recorded data from new stations, high-precision seismic response predictions can be realized, illustrating the adaptability and efficacy of the proposed approach in data-scarce conditions.