Comparing artificial neural networks with traditional ground-motion models for small magnitude earthquakes in Southern California

Alexis Klimasewski, Valerie J. Sahakian, & Amanda M. Thomas

Published August 13, 2019, SCEC Contribution #9496, 2019 SCEC Annual Meeting Poster #021

Earthquake ground motions are the superposition of source mechanics, wave propagation, and near site effects which cannot be fully known a priori. Statistically-based ground-motion models (GMMs) are by far the most widely-used method to estimate ground motions. GMMs typically regress for coefficients to parameterize the source, path, and site variables. Conversely, machine learning techniques, which are becoming increasingly popular, allow for fully data driven models without assuming a functional form. Machine learning ground-motion models could be useful for understanding regional behavior and in informing fully non-ergodic ground motion models. Here we compare these two approaches and investigate the performance and behavior of two methods of creating ground-motion models: a mixed-effects maximum-likelihood model with site correction term and an artificial neural net.

Our models estimate horizontal peak ground acceleration (PGA), and are created using 3,357 earthquakes of moment magnitude (M) 2.7-5.72 in Southern California. The events occur between 2010 and 2016, and are recorded on 16 ANZA and Caltech stations. The 52,340 records are split into 60% training, 20% validation, and 20% testing sets. We use the same splits for both methods to allow for a direct comparison. We create five-coefficient and six-coefficient mixed-effects models using the method of Sahakian et al. (2018). The five-coefficient model is a function of M and hypocentral distance (Rhyp), and the six-coefficient model includes a site term, either VS30 or κ0. We compare the five-coefficient model to a neural net with inputs of magnitude and distance. We choose the neural net architecture using the Akaike information criterion (AIC) and the standard deviation of residuals between observed and predicted PGA of the validation data. We evaluate the models using residuals between observed and predicted PGA of the test data.

We find that the five-coefficient mixed effects model performs better than the neural net model with only M and Rhyp inputs. Both six-coefficient mixed effects models perform similarly to the neural net GMPEs that include a site term. While the methods have similar performance, the scaling with distance and magnitude differ. Neural nets show promise in understanding region specific relationships between input parameters and ground motions.

Citation
Klimasewski, A., Sahakian, V. J., & Thomas, A. M. (2019, 08). Comparing artificial neural networks with traditional ground-motion models for small magnitude earthquakes in Southern California. Poster Presentation at 2019 SCEC Annual Meeting.

Related Projects & Working Groups
Ground Motions