[S-CG60] Driving Solid Earth Science through Machine Learning

[SCG60-P01] Improvement of the model architecture of the neural phase picker optimized for the JMA-unified catalog data

*Makoto Naoi¹ (1.Hokkaido University)

Keywords:JMA-unified catalog, Neural Phase Picker

Routine seismic observation data in Japan contains a large amount of waveforms and arrival time records. Naoi et al. (2024; EPS) retrained the deep learning model for arrival time reading (neural phase picker) PhaseNet (Zhu and Beroza 2019) using approximately 6 million waveforms with picking records that had been manually reviewed from 2013 to 2021 by JMA, and developed a model with high picking performance. In general, deep learning can achieve higher performance by using large models with high expressiveness as long as overfitting does not occur, but the learning curve obtained in the training of Naoi et al. (2024) showed that there were no signs of overfitting, even when using the PhaseNet architecture as is, or when using a model with double the number of channels in each convolution layer, suggesting that performance can be improved by using larger models. In this study, I attempted to extend the PhaseNet model by incorporating knowledge accumulated in the field of computer vision for creating larger, more computationally efficient models. As a result, I succeeded in creating a model that achieves a smaller loss value than would be difficult to achieve by simply increasing the number of channels or convolutional layers of the original PhaseNet.