Comparing Artificial Neural Networks with Traditional Ground-Motion Models for Small Magnitude Earthquake in Southern California

Session: Earthquake Source Parameters: Theory, Observations and Interpretations [Poster]
Type: Poster
Date: 4/30/2020
Time: 08:00 AM
Room: Ballroom
Description: 

Statistically-based ground-motion models (GMMs) are by far the most widely-used method to estimate ground-motions for seismic hazard. GMMs are the result of regressions to parameterize many predictive variables, with a user-specified relationship to ground-motion. Recently, machine learning techniques have become more popular for GMMs and allow for fully data driven models without assuming a functional form. The non-parameterization of machine learning GMMs could be useful for understanding regional seismological properties and behavior.

We compare the performance and behavior of these two approaches using a traditional mixed-effects maximum-likelihood (MEML) model with site correction term and an artificial neural net (ANN) trained on the same dataset. These models estimate horizontal peak ground acceleration (PGA) and are created with small magnitude earthquakes in Southern California, recorded on 16 seismic stations. The records are split into 60% training, 20% validation and 20% testing sets. This same partition is used for both methods, to allow for a direct comparison. The input parameters for both models include M and hypocentral distance (R_hyp) and some include a site parameter, either V_S30 or κ₀. We choose the ANN architecture from a suite of hidden layer and hidden unit sizes using the Akaike information criterion (AIC). We implement k-fold cross validation with 5 folds to stabilize model results. We evaluate both methods using residuals between observed and predicted PGA of the test data.

We find that with only M and R_hyp inputs, the MEML model performs better than the ANN model. Both MEML models with a site term perform similarly to the corresponding ANN GMMs; however, the scaling with distance and magnitude differ, with the ANN showing more specific models per site. Machine learning GMMs show promise in helping to understand region specific relationships between input parameters and ground motions.

Presenting Author: Alexis Klimasewski

Authors