Ground motion predictions based on soil properties using convolutional neural network

Earthquakes are among the most devastating natural disasters. In the past few decades, neural networks and other machine learning models have had significant success in geotechnical earthquake engineering. This paper studies the Convolutional Neural Network (CNN) model in predicting ground motion based on acceleration history measured in bedrock and soil properties. A CNN model is proposed; then, the model was trained and tested for 500 seismic events obtained from the Kiban Kyoshin Network (KiK-net, NIED 2016). When the predictions were compared with the baselines (actual measurements), the proposed CNN model presented good agreements.