[S-CG51] Driving Solid Earth Science through Machine Learning

[SCG51-P01] Machine learning approach to maximum magnitude estimates for seismic hazard

*Hiroe Miyake¹ (1.Earthquake Research Institute, University of Tokyo)

Maximum magnitude estimates are important for seismic hazard assessment. Toward time-dependent seismic hazard assessment, we explore the possibility of maximum magnitude estimates based on earthquake catalog and waveform analyses by machine learning approach. As for the earthquake catalog analyses, there are a lot of earthquake forecasting studies that count earthquake events to find laws of earthquakes. Traditional and statistical models such as the Omori formula, the Modified Omori formula, and the ETAS model, had a good performance on the aftershock forecasting but are sometimes not good enough for the mainshock forecasting. Liu, Miyake, and Tsuruoka (2019, AGU) considered using Recurrent Neural Networks to find the temporal relationship among earthquakes at different positions and to perform earthquake event forecasting. They used earthquake event data within 100 km depth in 2000-2011 from the JMA unified catalog, and sorted out one-day JMA magnitude probability distribution values of events number, magnitude average value, maximum magnitude value, and standard value, as one dataset per spatial grid to cover the catalog over Japan. Through supervised learning, they forecasted earthquake maximum magnitude based on Long-Short Term Memory Recurrent Neural Network prior to the 2011 Tohoku earthquake. As for the waveform analyses, Wu and Miyake (2019, AGU) examined the predictability of earthquake magnitude from the seismic nucleation phase via machine learning that may be helpful for realtime processing of earthquake early warning. Typical three models of seismic nucleation phase exist to explain how rupture grows and becomes an earthquake; self-similar model, pre-slip model, and cascade model. Since the self-similar model had been disproved, the controversy focuses on the nucleation model and the cascade model. The nucleation model indicates that earthquakes start from a weak pre-slip on the fault, then grow up. According to the hypothesis, the duration of the seismic nucleation phase is proportional to the event size of the earthquake. On the other hand, the cascade model indicates that the earthquake grows cascadingly and has no obvious nucleation phase. They tried to verify the hypothesis with machine learning approaches to provide some evidence. The NIED Hi-net data, the JMA unified hypocenter catalog, and the NIED seismic station information are used to create the dataset. They also designed methods to extract the nucleation phase among waveforms. Wave segments of the seismic nucleation phase are set as input, and earthquake magnitudes are set as an output in the machine learning process. MLP (multilayer perceptron), 1D CNN (Convolutional Neural Network), and KNN (k nearest neighbors) are implemented to compare their performances.