Ground Motion Model for Small-to-Moderate Potentially Induced Earthquakes using Machine Learning Algorithms
Description:
Ground motion models play vital roles in probabilistic seismic hazard analyses and seismic uncertainties. New ground motion models (GMM) are developed using non-parametric machine learning algorithms, including neural network, gradient boosting, support vector, k nearest neighbors, and random forest regression techniques. This paper evaluates the different machine learning models in predicting peak ground acceleration (PGA) and 21 spectral accelerations given the moment magnitude (Mw), rupture distance (Rrup) and average shear wave velocity of the upper 30 m of soil (Vs30). A database of 1336 ground motions (with small and moderate moment magnitude) ranging in magnitude from 2.8 to 5.8, recorded within a rupture distance range of 6–500 km in Central and Eastern North America, is used to train the algorithms. Linear regression-based models with predefined equations and coefficients are widely utilized. The requirement for predefined equations can restrict the use of complicated and nonlinear equations to improve performance. Compared to typical regression linear methods, the proposed GMM can increase the accuracy of GMM. Although the conventional regression model is more interpretable, machine learning can achieve a better result if enough training data is available. To evaluate the performance of different regression techniques in machine learning, mean square error (MSE), mean absolute error (MAE), explained variance score (Var), and the coefficient of determination (R2) are considered. Gradient boosting regression offers a better performance according to error metrics.
Furthermore, a machine learning weighted ensemble method is applied to develop the hybrid model. The weighted ensemble method is used to improve the GMM performance by combining the regression results of the algorithms. Based on the evaluation of regression models, the coefficient of determination (R2) increase in the hybrid model compared to different machine learning technique.
Session: USGS National Seismic Hazard Models: 2023 and Beyond [Poster]
Type: Poster
Date: 4/18/2023
Presentation Time: 08:00 AM (local time)
Presenting Author: Shahram Pezeshk
Student Presenter: No
Invited Presentation:
Authors
Najme Alidadi Corresponding Author nalidadi@memphis.edu University of Memphis |
Shahram Pezeshk Presenting Author spezeshk@memphis.edu University of Memphis |
|
|
|
|
|
|
|
Ground Motion Model for Small-to-Moderate Potentially Induced Earthquakes using Machine Learning Algorithms
Category
USGS National Seismic Hazard Models: 2023 and Beyond