Poster #225, Ground Motions

Deep Learning for Site Response Estimation from Geotechnical Array Data

Kim B. Olsen, & Daniel Roten
Poster Image: 

Poster Presentation

2020 SCEC Annual Meeting, Poster #225, SCEC Contribution #10585 VIEW PDF
Ground motions recorded on vertical arrays show that theoretical methods for site response prediction frequently fail to reproduce the observed surface-to-borehole amplification function due to modeling simplifications. We propose to improve site response prediction by harnessing the large amount of strong motion data collected on geotechnical (borehole) arrays as a training dataset for a deep learning algorithm, with the goal to predict site response without relying on simplifying assumptions or proxies.

Our approach is based on an artificial neural network (ANN) consisting of multiple layers, where the input layer contains the discretized shear-wave velocity profile, frequ...
ency of amplification, as well as earthquake characteristics including magnitude, distance and shaking duration. The output layer consists of a single neuron representing the amplification at the selected frequency. Experiments with 600 real KiK-net soil profiles show that a properly regularized ANN with seven hidden layers is able to learn the theoretical SH amplification functions, and predict the site response for profiles that were not part of the training set. We train the network using a large number of observed amplifications from vertical arrays located in California (CSMIP) and Japan (KiK-net). We find that predicting observed transfer functions from unsmoothed velocity profiles requires more hidden layers than predicting theoretical amplifications from simplified soil profiles. We need at least nine hidden layers to obtain low bias while training the network with transfer functions observed at Japanese and Californian vertical arrays.

We experiment with different network architectures and regularization strategies to improve the predictions and to avoid overfitting. The accuracy of the network is evaluated against traditional (theoretical) site response assessment techniques for randomly chosen sites and earthquakes which were not part of the training dataset.