Taiwan is located in Circum-Pacific Belt and has been suffering from frequent earthquakes. The damage caused by large earthquakes is devastating and therefore, seismological study is one of the major academic scientific researches in Taiwan. The precise measurement of the P and S waves will help seismologists to calculate the distance from the epicenter of an earthquake. To analyze seismic data could conduce to better understanding of warning signs of an earthquake. And combining with data of precise earthquake waves will improve accuracy and speed of early warning systems.
The interpretation of seismic wave motions takes a lot of efforts, especially the P and S wave records from multiple seismographs, which takes time for experienced analysts or engineers to pick and analyze the data for further calculation and explanation. Compared with the P waves which are easier to analyze, the standard for S waves will have certain variations among different seismologists.
Taiwan government has deployed more seismograph stations in the island to gather huge data than before. Dealing with big quantities of data under the conditions of time constraint and analysts shortage, it is an intense pressure for public sectors including Taiwan Central Weather Bureau (CWB). When earthquake happens, the immense data travels from different stations. The potential of using automation systems to assist or even replace the traditional manual operation is becoming an important trend. The Artificial Intelligence (AI) technology is a new star to assist solving the needs.
The ARRU seismic phase picker project of Earth Sciences Department at the National Cheng Kung University (NCKU) is aiming to use Deep Learning models to select right waves from massive data. The project leader is Will Liao, who is PhD student in his first year. In this interview, he highly appreciates the effort and time devoted by Central Weather Bureau and research colleagues in the community to work on keeping the seismic data of earthquakes over the past decades. These efforts also allow the following researchers to continue the works as if standing on the shoulders of these giants.
Liao explains the detection rate of traditional automatic analysis method for P-Wave is around 75% while there is only 30% for S-Wave because of refraction or interference from P-Wave. To improve this phenomenon, the analysts and engineers take lots of time to correct the errors by re-checking and re-annotating in order to obtain a reasonable digit. The traditional automation method is to solve the issues, but it's not handy and instead cause more errors because of high variation of S-Wave. Some will even get more errors while multiple earthquakes happen in a short period of time; therefore the results of the automation method were far behind compared with human working results.
This AI-related project will include the seismic data from Central Weather Bureau and other research organizations. There are about 30 to 35 gigabyte data files reserving for use in the AI training. Those big data sets require huge GPU computing power. Initially, Liao used GeForce RTX 2080Ti GPU accelerator card in NCKU laboratory which took two weeks for just running the AI training, but things have been improved. Through participating in the AI start-ups TWCC STAR TREK" hosted by National Center for High-performance Computing (NCHC), this project gets support from Taiwan Computing Cloud (TWCC) platform, and now the AI training takes only 4 or 5 days which is inspiring.
The current stage is using TWCC's single GPU accelerator card for the first trials. In the next step, it will enable multiple GPU accelerator cards for AI training by modifying the Python codes. The target is to train more than 100 AI models. So far, with about two weeks trail experience, Liao has seen preliminary results. In the coming stages, it's expected to achieve more than 5-times computing performance with the help of V100 GPU and more memory capacity.
The preliminary results using current effective AI models have fairly improved of P-wave and S-wave detection rate which ramping up to 98% and 90% respectively. Compared with previous results of 75% and 30%, the outcomes are very impressive. Combined with the conventional physics models, it is showing the convincing results and has better usage in the study of earthquake models which is associated with seismogenic structures. This research has been submitted to Q1 ranking journals in October 2020. The future goal for this project will aim to bring AI inference system to Edge Computing platforms and deploy to different sites for helping Taiwan Central Weather Bureau to increase the accuracy and speed of earthquake detection.
Will Liao, PhD student at National Cheng Kung University
DIGITIMES' editorial team was not involved in the creation or production of this content. Companies looking to contribute commercial news or press releases are welcome to contact us.
