透過深度學習判斷地震波極性及其在決定震源機制解中之應用

Abstract

台灣的地震頻繁且常伴隨災害發生。震源機制解提供斷層機制和應力場資訊，有助於理解斷層破裂過程及評估緊急災害應變。因此，除了震源、發震時間、位置及規模等基本資訊外，迅速獲取震源機制解對及災害減輕也至關重要。傳統決定震源機制解的方法主要分為質心地震矩逆推法(Centroid Moment Tensor, CMT)和P波初動法兩類。大地震的破裂機制往往複雜且與初始破裂行為不一致；CMT方法提供的震源機制解能反映震源破裂的完整過程，而使用P波初動法所獲得的震源機制解則反映震源初期的破裂行為。雖然兩方法得到的解代表著不同的物理意義，但這些方法通常需要幾分鐘到數十分鐘才能獲得震源機制解。 近年來，深度學習的迅速興起在地震學取得重大突破，例如地震偵測、地震相位拾取和P波初動極性確定等，與傳統方法相比提供了更高效且自動化的數據分析能力。本研究提出一種基於深度學習的方法，用於識別地震波的極性，並應用於決定震源機制解，目標旨在短時間內獲得與CMT相近的解。本研究以台灣P-alert測網2012年至2024年4月間規模大於4的島內事件，使用短時間窗的三分量加速度紀錄作為訓練資料，將地震波極性以GCMT目錄為參考解標記為上動、下動與不確定三類別進行訓練。透過交叉驗證，模型以平均97.3%的準確率區分上下動之極性。在對台灣各區域10起獨立地震事件的進一步性能測試中，模型預測結果成功應用於所有測試事件之震源機制解解算，整體準確率達82.21%，表明模型具備良好的泛化能力。同時，本研究也透過Kagan角量化測試結果與GCMT、CWA、AutoBATS目錄中震源機制解的相似程度。結果顯示，所有事件的Kagan角均低於50度，其中與參考解GCMT目錄更是低於35度。這顯示本研究提出的模型能準確判斷地震波的極性，並與現有的震源機制解目錄具有高度的相似性。此外，在模擬P波到時不確定性的時間窗平移測試中，模型在±0.2秒的平移範圍內也展現出穩定的極性辨識能力，顯示其對於P波到時之精準性的容錯性與適應性良好，進一步凸顯本研究模型在實際應用中的可靠性與發展潛力。

In Taiwan, earthquakes are common and often lead to disasters. Focal mechanism solutions reveal the faulting mechanism and stress field, contributing to understanding the rupture process and assessing seismic hazards. Therefore, in addition to essential earthquake information such as origin time, location, and magnitude, rapidly obtaining focal mechanism solutions is crucial for emergency response and damage mitigation. The inversion methods of focal mechanism solutions can be sorted into two main categories: Centroid Moment Tensor (CMT) and P-wave first-motion. The rupture mechanism of large earthquakes often differs from their initial rupture. Focal mechanism solutions derived from CMT represent the overall focal rupture process, while those obtained through the P-wave first-motion method capture the initial focal rupture behavior. Although the physical meaning of focal mechanism solutions from the two methods is different, these conventional approaches generally require several to tens of minutes to resolve the solutions. In recent years, the rise of deep learning has led to significant breakthroughs in seismology, such as earthquake detection, seismic phase picking and P-wave first motion polarity determination. Compared to conventional methods, deep learning offers a more efficient and automated approach to data analysis. In this study, we proposed a deep learning approach for identifying the polarity of seismic waves and applied it to determine the focal mechanisms. Our goal is to enhance the efficiency of obtaining focal mechanisms. The data used in this study were collected from the Taiwan P-alert network, covering inland earthquakes with magnitudes greater than 4 that occurred between 2012 and April 2024. Unlike previous studies, our model was trained using three-component acceleration signals within a short time window, with waveform polarities labeled as up, down, and uncertain based on the GCMT catalog. Through cross-validation, the model achieved an average accuracy of 97.3% in distinguishing the polarities. In a further performance test involving 10 independent earthquake events from different regions in Taiwan, the model’s predictions were successfully applied to carry out all the focal mechanism solutions. The evaluation resulted in an overall accuracy of 82.21%, indicating strong generalization capability of the model. Additionally, we used the Kagan angle to quantify the similarity between our results and focal mechanism solutions from GCMT, CWA, and AutoBATS CMT catalogs. The results show that the Kagan angles between our solutions and three reference catalogs are below 50°. Particularly, those compared to the GCMT reference solutions are consistently below 35°. This demonstrates that the deep learning model in this study can accurately determine waveform polarity, and the focal mechanism solutions derived from our approach exhibit a high degree of similarity with those in existing public catalogs. Furthermore, in the time-shifting test simulating uncertainty in P- arrival time, our model maintains consistent polarity classification performance within a ± 0.2 second shifting range. This highlights the model’s robustness against P-arrival time uncertainty and indicates its reliability and potential for practical applications in rapid and reliable focal mechanism determination.