花蓮地區位處歐亞板塊與菲律賓海板塊碰撞縫合帶，地質構造複雜，近年頻繁發生破壞性地震。本研究旨在建立花蓮地區高解析度三維速度與密度構造，並以2024年4月3日花蓮地震為核心事件，透過不同時間段之地震層析成像，探討斷層帶流體行為與震後應力調整機制。在資料方面，使用2012–2020年中央氣象署地震目錄與體波到時資料，並整合中研院、中油、內政部之重力觀測資料，作為三維速度與密度基礎模型之輸入；同時運用EQTransformer與GaMMA演算法重建2024年花蓮地震序列之AI地震目錄。在分析架構上，透過地震—重力序列反演建構兼顧深淺構造之基礎模型，並以其作為初始模型，對震前、震後地震目錄分別進行獨立反演，以追蹤速度構造之時間變化。基礎速度模型清晰辨識出北縱谷地區的中央山脈高速變質岩體和縱谷低速沉積帶的側向不連續，並於15–25公里深處辨識出中央山脈斷層（CRF）的深部延伸。AI地震目錄偵測事件數顯著優於氣象署目錄，原始時間殘差之標準差為0.743秒，三維重定位誤差於水平方向約600–700公尺，垂直方向約741公尺。速度構造隨時間變化方面，震前縱谷斷層（LVF）附近的高Vp/Vs特徵指示流體飽和的孕震環境；震後北段（A、B 剖面）LVF深部低速異常出現速度回升，推測為同震滲透率提升所驅動之孔隙壓力釋放與裂隙閉合所致；南段（C、D剖面）CRF深部則呈現同震岩石破壞引發的速度下降，並於CRF–LVF交界處出現局部速度升高，顯示應力轉移與流體重分配的疊加效應。上述南北不對稱的速度時間變化，反映主震破裂由LVF深段起始、觸發淺段CRF的時空演化過程，並指示CRF–LVF幾何截斷關係在控制此次地震序列發震機制與震後流體重分配上扮演關鍵角色。本研究提供花蓮地區首個結合地震–重力聯合反演與AI地震目錄的四維速度–密度構造框架，為後續斷層帶流體動態與震後應力演化等相關模擬提供初始約束。;The Hualien region lies along the collisional suture between the Eurasian and Philippine Sea plates， where structural complexity has produced frequent destructive earthquakes in recent years. This study aims to construct a high-resolution three-dimensional velocity and density model of the Hualien region， and uses the 3 April 2024 Hualien earthquake as the central event to investigate fault-zone fluid behavior and post-seismic stress adjustment through tomographic imaging across different time windows. For input data， we use the 2012–2020 Central Weather Administration （CWA） earthquake catalog and body-wave arrival times， together with gravity observations integrated from Academia Sinica， CPC Corporation， and the Ministry of the Interior， as the basis for constructing a 3-D velocity and density baseline model. In parallel， EQTransformer and the GaMMA algorithm are employed for automatic phase picking and association to reconstruct an AI-derived earthquake catalog for the 2024 Hualien sequence. Methodologically， a seismic–gravity sequential inversion framework is applied to build a baseline model that jointly resolves shallow and deep structures; using this as the initial model， the pre- and post-mainshock catalogs are inverted independently to track temporal changes in the velocity structure. The baseline model clearly delineates the lateral contrast between the high-velocity metamorphic basement of the Central Range and the low-velocity sedimentary fill of the Longitudinal Valley in the northern Longitudinal Valley region， and resolves the deep extension of the Central Range Fault （CRF） at depths of 15–25 km. The AI catalog detects substantially more events than the CWA catalog; the standard deviation of arrival-time residuals relative to manual picks is 0.743 s， and 3-D relocation uncertainties are approximately 600–700 m horizontally and ~741 m vertically. Regarding temporal velocity variations， the high Vp/Vs anomaly near the LVF prior to the mainshock indicates a fluid-saturated seismogenic environment. After the mainshock， in the northern profiles（AA′ and BB′） the deep low-velocity anomaly along the LVF shows velocity recovery， tentatively attributed to coseismic permeability enhancement that drove pore-pressure release and crack closure. In the southern profiles（CC′ and DD′）， the deep CRF exhibits velocity decreases consistent with coseismic rock damage， while localized velocity increases appear near the CRF–LVF junction， suggesting the combined effects of stress transfer and fluid redistribution. The north–south asymmetry in temporal velocity variations reflects a rupture history in which the mainshock initiated on the deep LVF segment and subsequently triggered the shallow CRF， indicating that the geometric truncation relationship between the CRF and LVF plays a key role in controlling the nucleation mechanism and post-seismic fluid redistribution of this earthquake sequence. This study provides the first four-dimensional velocity–density framework for the Hualien region integrating seismic–gravity sequential inversion with an AI-derived earthquake catalog， offering initial constraints for future quantitative simulations of fault-zone fluid dynamics and post-seismic stress evolution.
