A data-driven approach for spectrum-matched earthquake ground motions with physics-informed neural networks

Abstract

This study presents a novel data-driven approach for generating spectrum-matched earthquake ground motions using physics-informed neural networks (PINNs). The methodology leverages real recorded earthquake data and employs singular value decomposition for dimensionality reduction, enabling the extraction of eigen motions that capture correlated temporal patterns. By combining PINNs with these eigen motions, spectrum matching is achieved with clear physical interpretability. The generated motions balance conventional linear scaling and spectrum matching, with the degree of matching dependent on the input motions, while retaining the realistic non-stationary features inherent in the input data. The adequacy of the post-matched motions is evaluated through various measures and incremental dynamic analysis to identify any potential biases introduced by the spectral matching process. The findings indicate that, despite some deviations in spectral shape, the overall performance of the spectrum-matched motions remains acceptable, without introducing significant bias.

Keywords

• Earthquake ground motion
• spectrum matching
• Physics-informed neural networks
• Seismic design
• Data-driven engineering