← Back to Sessions

WITHDRAWN Applications of Machine Learning to Earthquake Early Warning and Ground Motion Prediction Equations

Description: 

WITHDRAWN In this study, we applied machine learning approaches to predict possible strong ground motions for earthquake early warning (EEW) and ground motion prediction equations (GMPEs). Regarding EEW, many AI-based algorithms have been proposed for on-site warning. In contrast, the waveforms' spatial characteristics are seldom implemented into the algorithm, with one exception being Münchmeyer et al.’s Transformer Earthquake Alert Model (TEAM) (2021). In our study, we applied the TEAM model to the Taiwan region by implementing the strong-motion waveforms obtained by the Taiwan Strong Motion Instrumentation Program (TSMIP) network. We trained the TEAM model using features of the traces within three seconds after P-wave arrivals that the convolution neural network had extracted and the transformer model linked to station locations. We further expanded our model to predict not only peak ground acceleration (PGA) but also peak ground velocity (PGV) to provide a wider application to end users.

To establish a set of AI-based GMPEs, we compared the performance of the models considering Support Vector Regression, Random Forest Regression, Gradient Boosting Regression, and XGBoost Regression and reduced the prediction error by taking advantage of ensemble learning. We applied this approach to build GMPEs for crustal earthquakes in Taiwan by implementing the strong-motion records from the TSMIP network. Comparing our results with observations, we found our GMPEs underestimated the ground shakings for large events due to insufficient observations. To solve the imbalanced numbers of events for some parameters (e.g., magnitudes, rupture distance, and site amplification factor), we implemented GAN, SMOTR, and Gaussian noise methods to synthesize artificial data for training to improve model performance. Furthermore, we used the grid search method to find the best solution for the hyperparameters and collect total combinations into MySQL to identify differences. We presented the GMPEs in the forms of not only PGA and PGV, but also various periods of spectral acceleration (SA).

Session: Opportunities and Challenges for Machine Learning Applications in Seismology

Type: Oral

Date: 4/19/2023

Presentation Time: 05:15 PM (local time)

Presenting Author: Chung-Han Chan

Student Presenter: No

Invited Presentation: 

Authors