Leveraging physics-informed neural networks in geotechnical earthquake engineering: An assessment on seismic site response analyses

Abstract

The primary objective of this study is to assess the potential of physics-informed neural networks (PINNs) for seismic site response analyses (SRA). PINNs constitute a novel computational paradigm that combines physical principles with data-driven methods to solve differential equations. Despite the growing exploration of machine learning and deep learning in geotechnical earthquake engineering, the integration of PINNs remains limited. The study first addresses key challenges in applying PINNs to SRAs. In particular, the broad range of frequencies in ground motion recordings, which complicates the training process, and neural network architectural issues are discussed. Fourier feature embedding, a relatively new technique in image processing, learning rate adjustment, a tailored training strategy, and a PINN architecture are proposed to address the identified challenges. The proposed framework is evaluated by comparing SRA results from the implemented PINN and traditional numerical techniques, considering different ground motions and soil systems. The results of the proposed PINN and numerical techniques are identical, highlighting the robustness of the proposed framework. The encouraging results suggest there is significant potential for PINNs in general geotechnical earthquake engineering applications, which is also discussed.