SCEC Award Number 21160 View PDF
Proposal Category Individual Proposal (Integration and Theory)
Proposal Title Deep Learning for Amplitude-dependent Prediction of Surface-to-Borehole Site Response
Investigator(s)
Name Organization
Daniel Roten San Diego State University Kim Olsen San Diego State University
Other Participants
SCEC Priorities 4a, 4c, 4d SCEC Groups GM, CS, Seismology
Report Due Date 03/15/2022 Date Report Submitted 11/14/2024
Project Abstract
 We investigate the use of deep learning for the prediction of surface-to-borehole Fourier amplification functions (AFs) from discretized shear-wave velocity profiles for individual earthquakes. This project builds on a previous study that used fully connected and convolutional neural networks (NNs) to predict mean AFs during weak ground motion, and aims to capture path- and amplitude-dependent site effects, in particular nonlinear site response.
We design a mixed-input neural network that accepts the velocity profile as sequence data and earthquake parameters including magnitude, epicentral distance, and depth as numerical parameters, and predicts the AF for a range of predefined frequencies.
The network learns to reproduce the recorded amplification function from the training dataset and reasonably predicts AFs for randomly selected test sites. In particular, the network captures nonlinear effects present in the training data and predicts decreasing site amplification with increasing earthquake magnitude, in particular at higher frequencies, at near-fault locations.
Intellectual Merit This proposal focuses on the dependency of strong ground motions on nonlinearities (research priority 4), in particular shallow crust nonlinearity (4a). The deep learning model used in this proposal can be used as an alternative to empirical nonlinear site correction factors used in physics-based ground motion maps (4d, 2b). In contrast to established empirical site response prediction models, the NN can use high-resolution near-surface information (4b).
Broader Impacts The development of state-of-the-art ground motion models and earthquake hazard maps represents a fundamental aspect of SCEC's broader impact and public outreach activities. This modernization is driven by improved physical modeling, denser seismic networks, and the advent of deep learning techniques. This work contributes with a novel application of deep learning that could be used to incorporate nonlinear effects in ground motion and seismic hazard maps.
Exemplary Figure Figure 4: Recorded and predicted AFs at the test site KMMH14. AFs from weak motion records are shown in gray. The green line shows the AF during the 2008 M 7.2 Iwate–Miyagi Nairiku earthquake.
Linked Publications

Add missing publication or edit citation shown. Enter the SCEC project ID to link publication.