[S-CG50] Driving Solid Earth Science through Machine Learning

[SCG50-08] Attention-based Machine Learning Model for Magnitude Estimation

*JI ZHANG¹, Aitaro Kato¹, Huiyu Zhu² (1.Earthquake Research Institute, the University of Tokyo, 2.University of Science and Technology of China)

Keywords:Machine Learning, Magnitude estimation

Rapid and reliable earthquake magnitude estimation is important for earthquake early warning (EEW), particularly during the critical early stages of event detection. Traditional magnitude estimation methods rely on complete waveform records, including earthquake epicenter distance and waveform amplitude, leading to delays in magnitude assessment. Machine learning techniques offer a promising avenue for capturing nonlinear relationships within seismic data, enhancing information extraction and timeliness in earthquake magnitude estimation. In this study, we introduce an Attention-based Machine Learning model for Magnitude Estimation (AMAEnet) tailored for EEW applications. Our approach enables magnitude prediction within a minimum of 1 second of the P-wave arrival, without constraints on input signal size. Utilizing two independent open-source datasets for training and testing, our results demonstrate the efficacy of our method in accurately predicting earthquake magnitudes. We systematically explore the impact of network architecture, loss functions, and signal length on prediction performance. Evaluations encompass various network depths (3 to 7 layers) and convolution kernel sizes (3, 5, and 7). Notably, our findings reveal that a network with a depth of 4 and a convolution kernel size of 5 yields optimal prediction accuracy. Additionally, we assess different loss functions including mean square error (MSE), mean absolute error (MAE), weighted loss function, and loss function with standard deviation constraint, highlighting MSE as the most effective. Furthermore, we investigate the influence of input signal length and the integration of attention mechanisms. The results indicate that a signal length of 6 seconds, with noise and signal lengths having equal durations (3 seconds each), optimizes model prediction accuracy. The attention mechanism facilitates the identification of the first motion while providing insights into the network's focus areas, thereby establishing a relationship between waveform characteristics and earthquake magnitude. In conclusion, our study underscores the potential of machine learning-based magnitude estimation in EEW systems, offering novel opportunities to mitigate natural disaster impacts, minimize casualties, and safeguard lives and property.