利用機器學習量測表面波頻散探測SAA深度範圍

Abstract

本研究旨在探索台灣地區平行應力方向的非均向性（Stress-Aligned Anisotropy ,SAA）的深度範圍。近期利用尾波干涉法研究的結果顯示，台灣的海岸平原近地表SAA與利用本地的淺層地震進行的剪切波分裂（Shear wave splitting, SWS）研究結果一致，然而，這與使用深震的SKS的SWS研究，以及從噪聲衍生的寬頻（4-20秒）雷利波導出的淺層地殼Vs非均向性模型並不完全吻合。這表明SAA很可能僅侷限於淺層地殼，儘管可以合理推估岩層中微裂縫將會因為圍壓隨著深度不斷增加而閉合，但台灣SAA存在的深度範圍仍有待探索。 本研究使用周遭噪訊成像法，利用臺灣西南部密集測站序列資料，建構該區域的三維速度構造模型。為了解析淺層構造，我們計算了測站對的交互相關函數，並聚焦於短週期（2-7秒）的雷利波。為了準確且有效地進行量測短週期的頻散曲線，我們採用了基於Yang et al.(2022)提出的機器學習演算法，並結合Liao et al.(2021)開發的循環殘差U型網絡（R2U-Net）進行訓練。模型訓練數據包括來自臺灣、日本和紐西蘭南島等不同地區計算的頻散圖，從而建立了一個更加高效且客觀的模型，並成功應用於本研究短週期表面波的頻散測量。 通過獲得的頻散數據，本研究採用以小波為基底函數的多尺度反演技術導出剪力波與方位非均向性的三維模型。結果顯示，密集的短週期表面波可以有效地解析淺層的SAA構造。在此三維模型中，臺灣西部海岸平原區的SAA深度範圍可延伸至6公里以下，變形前緣以東則漸轉為受構造主導的非均向性；結果也顯示SAA由淺層的西北-東南向隨著深度增加逐漸轉至東西向，這可能歸因於在該深度造山運動前即存在之東西走向的正斷層系統的影響。

In this study, we aim to explore the depth range of stress-aligned anisotropy (SAA) in Taiwan. Our recent works using coda-interferometry (SWS) have shown that the near-surface SAA is consistent with shear-wave splitting studies employing local earthquakes. However, it contrasts with SWS studies using deep phase (SKS) and the shallow crustal Vs anisotropy model derived from noise-derived broad-band (4-20 s) Rayleigh waves. This suggests that SAA is likely confined to the uppermost crust. Despite micro-cracks assumed to be fully closed with increasing ambient stress at depths, the depth range of the SAA mechanism remains unclear. Our approach involves noise tomography using short-period (2-7s) Rayleigh waves enhanced by multicomponent stacking technique. To measure the dispersion of the isolated fundamental mode Rayleigh waves accurately and effectively, we employ a modified machine learning algorithm bases on the algorithm proposed by Yang et al.(2022) We employ the Recurrent-Residual U-Net (R2U-Net) developed by Liao et al.(2021)for training. The model training data consist of dispersion diagrams from CCFs derived in various regions, including Taiwan, Japan, and the South Island of New Zealand. These data have enabled the development of a more efficient and objective model, which has been successfully applied to the study of short-period surface waves. With the obtained dispersion data, we apply the wavelet-based multi-scale inversion technique to derive 3D models of Vs and Vs anisotropy .The results indicate that the depth of the SAA extends down to 6 kilometers and likely deeper. The area west of the deformation front is dominated by SAA, while the area to the east is influenced by Orogeny-Parallel Anisotropy (OPA). The direction of the SAA shifts from northwest-southeast to east-west with increasing depth, which is likely influenced by pre-existing normal fault systems oriented in the east-west direction.