Spatiotemporal Stress Changes During the 2024 M7.6 Noto Peninsula Earthquake Sequence Revealed From Polarity‐Based Focal Mechanisms via Siamese Neural Network

Abstract Detailed stress field analyses are vital for understanding earthquake nucleation and assessing potential seismic hazards. The M7.6 Noto Peninsula earthquake sequence, which recorded a significant number of small‐to‐moderate earthquakes, offers an excellent opportunity to study the detailed spatiotemporal stress redistributions. Deep learning has demonstrated its effectiveness in determining polarity‐based focal mechanisms for small‐to‐moderate earthquakes. In this study, we first propose a Siamese neural network (SNN) based on similarity learning and evaluate its prediction performance against the conventional Convolutional neural network (CNN) on a test data set. SNN model shows slightly improved performance than the CNN model in identifying more polarities. We then apply both the CNN and SNN models to analyze the M7.6 Noto Peninsula data set, obtaining four times as many focal mechanisms as the AQUA catalog. All three sets of focal mechanisms (from the SNN, CNN, and AQUA) are used to invert the spatiotemporal stress changes. Stress inversion results show that, temporally, the azimuthal rotation in S 1 caused by the mainshock is about 4.4° counterclockwise, which is relatively minor. Spatially, all three sets of stress results consistently reveal azimuthal rotations in S 1 of approximately 26°–35.7° across different subareas with estimation errors of 5.1° and 9.7°, respectively, suggesting significant spatial stress heterogeneity. By comparing with the mapped active faults, we find that the spatial azimuthal rotations in S 1 are generally coupled with the variation in fault orientations. Our findings indicate that, in addition to the complexity of fault geometry, localized stress heterogeneities may also contribute to the growth of complex earthquake ruptures.