MEANet: Magnitude estimation via physics-based features time series, an attention mechanism, and neural networks

Abstract

The traditional magnitude estimation method, which establishes a linear relationship between a single warning parameter and the magnitude, exhibits considerable scatter and underestimation. In addition, the extraction of features from raw waveforms by a deep learning network is a black box. To provide a more robust magnitude estimation and to construct a deep learning network with an interpretable input, in light of deep learning and earthquake rupture physics, we have established a magnitude estimation network model (MEANet) via the physics-based features time series, an attention mechanism, and neural networks. We use events with 4 ≤ M ≤ 7.5 that occur in Japan and the Sichuan-Yunnan region, China, to train and validate MEANet, and then use MEANet to test additional events. Our results find that MEANet has a more robust magnitude estimation than the traditional τ c and P d methods, with a standard deviation of error of ±0.25 magnitude units at a single station with a 3 s P-wave time window. Within 10 s after the first station is triggered, based on the weighted average of the triggered stations, MEANet provides robust magnitude estimation without underestimation for events with 4 ≤ M ≤ 7.5. Our finding implies that the final magnitude is to some degree deterministic by the combination of deep learning and physics-based features. Meanwhile, MEANet might have potential for earthquake early warning.

Combining Convolutional Neural Network with Physics- Based Features in Shallow and Intermediate-depth Earthquake Discrimination

“…In addition, they may implicitly pay close attention to features such as event location or timing while ignoring truly salient components of the seismic waveform, which may lead to a decrease in accuracy in practical applications. To overcome this problem, recent studies have shown that combining deep learning with physics parameter-based features can improve the performance of the deep learning model (Kong et al, 2022;Song et al, 2023).…”