← Back to Sessions

Pisgan: Physics-Informed Seismic Waveform Generator Trained With a Large-Scale Seismic Benchmark Dataset of China

Description: 

Every year, millions of earthquakes occur around the world. Some of the large ones can cause serious hazard to constructions and facilities. The prediction of ground motion in specific locations can help mitigate the seismic risk. Traditional physics-based ground motion modeling methods suffer the problems of model misspecification, high computational cost, and uncertainty quantification. In this study, we introduced PISGAN – Physics-informed Seismic Generative Adversarial Networks – as an alternative simulator to generate realistic seismic waveform. The PISGAN was implemented with the architecture of conditional deep convolutional GAN. We embedded physical information, such as epicenter distance, azimuth angle, and velocity model, as conditions to both the generator and the discriminator of our PISGAN. The ability of the PISGAN was first validated by synthetic tests. Then, the PISGAN was trained with data from DiTing, a large-scale Chinese seismic benchmark dataset. The trained PISGAN can generate realistic seismic waveforms with “Chinese characteristics” that can fool human experts and models, including traditional models and neural networks.

Session: Opportunities and Challenges for Machine Learning Applications in Seismology

Type: Oral

Date: 4/19/2023

Presentation Time: 04:30 PM (local time)

Presenting Author: Borui Kang

Student Presenter: Yes

Invited Presentation: 

Authors