← Back to Sessions

Ground Motion Models: Comparison Between Traditional Regression-based Techniques and Machine Learning Approaches

Description: 

This study aims at a comparison between traditional methods for predicting ground motion Intensity Measures (IMs), usually based on linear regression, and machine learning techniques, which are widely used today for different purposes thanks to the increasing availability of data and computing performance. Typically, Ground Motion Models (GMMs) are empirical equations to estimate IMs as a function of several variables, generally related to the seismic source (e.g. magnitude), source-to-site distance (e.g. Joyner-Boore distance), and local site conditions (e.g. site categories or VS30). An alternative could be the use of machine learning (ML) algorithms, non-parametric models that are fully trained by data. We investigate the efficiency of different ML algorithms to identify the most suitable for predicting ground-motion using the data set used by Lanzano et al (2019) to derive the most recent ground-motion model for Italy. We explore different Matlab© ML algorithms and we split the dataset randomly into two parts, 70% as training data and 30% as test data, such that each class is correctly represented in the resulting subsets. We find out that the Gaussian Process Regression (GPR) significantly reduces the standard deviation associated with the predictions. The predictions by the GPR are compared with the Lanzano et al. (2019) model, in order to quantify the differences in terms of standard deviation, which is broken down into between-event, between-station, and event- and site-corrected components, implemented as random effects.

The differences between the two approaches are maximized when the variable sampling is poor since the ML approach tends to reproduce, with small uncertainty, the few observations available, which could be extremes. Reliable predictions can be obtained when the combination of all predictor variables is well sampled by high volumes of data (e.g. 1000 observations or more). At present times this goal can be achieved with worldwide data sets.

Session: Opportunities and Challenges for Machine Learning Applications in Seismology

Type: Oral

Date: 4/19/2023

Presentation Time: 05:00 PM (local time)

Presenting Author: Lucia Luzi

Student Presenter: No

Invited Presentation: 

Authors