Predicting peak ground acceleration using the ConvMixer network

The level of ground shaking, as determined by the peak ground acceleration (PGA), can be used to analyze seismic hazard at a certain location and is crucial for constructing earthquake-resistant structures. Predicting the PGA immediately after an earthquake occurs allows for the issuing of a warning by an earthquake early warning system. In this study, we propose a deep learning model, ConvMixer, to predict the PGA recorded by weak-motion velocity seismometers in Japan. We use 5-s three-component seismograms, from 2 s before until 3 s after the P-wave arrival time of the earthquake. Our dataset comprised more than 50,000 single-station waveforms recorded by 10 seismic stations in the K-NET, Kiki-NET, and Hi-Net networks between 2004 and 2023. The proposed ConvMixer is a patch-based model that extracts global features from input seismic data and predicts the PGA of an earthquake by combining depth and pointwise convolutions. The proposed ConvMixer network had a mean absolute error of 2.143 when applied to the test set and outperformed benchmark deep learning models. In addition, the proposed ConvMixer demonstrated the ability to predict the PGA at the corresponding station site based on 1-second waveforms obtained immediately after the arrival time of the P-wave.