2004 年艾利颱風降下巨量雨水，誘發新竹五峰土場發生百萬立方公尺級之深層大崩塌。然而，降雨引致之邊坡大變形破壞與長距離流動機制常難以預測。針對此類災害，傳統數值分析常遭遇網格扭曲或無法模擬流動堆積之技術瓶頸。為突破此限制，本研究引入基於雙相單點物質點法 (MPM) 之 Anura3D 軟體作為分析工具。本研究首先於二維模型重現大型降雨入滲物理試驗，以地表位移與內部軸差應變為比對指標，證實該方法具備捕捉塊體滑動與剪裂帶演化歷程之能力。同時透過參數敏感性分析，確立針對鬆散崩積層引入「負剪脹角」以模擬「剪縮效應」，為合理反映土體受剪切時激發超額孔隙水壓及轉化為泥流狀態之重要力學因素。 於三維空間模擬方面，本研究整合災前災後數值地表模型 (DSM) 之高程差分析 (DoD) 與地質鑽探資料，建構出具備土場深層滑動面特徵之三維幾何模型。考量軟體於動態水文邊界設定之限制，本研究直接將降雨致災結果轉化為模型之初始高地下水位，進行三維大變形流固耦合分析。模擬結果顯示，在高地下水位之驅動下，鬆散土體受剪切所引發之剪縮效應會激發局部超額孔隙水壓，並使有效應力下降，進而產生土石流態之大規模流動。最後，透過混淆矩陣 (Confusion Matrix) 進行空間預測量化評估，本研究確立之基準參數模型其精確率達 58.36%、召回率達 68.71%，交併比 (IoU) 為 46.11%。此量化結果顯示，本模型之預測結果與實際災後堆積形貌具備合理之空間對應關係，可為未來類似崩塌之減災防護提供參考。;In 2004， Typhoon Aere triggered a deep-seated， million-cubic-meter landslide in Tuchang， Hsinchu. Traditional numerical analyses of such large-deformation landslides often face technical bottlenecks like mesh distortion. To overcome these limitations， this study employed Anura3D， based on the two-phase single-point Material Point Method (MPM). Initially， a 2D MPM model successfully reproduced a large-scale rainfall infiltration test， verifying its capability to capture block sliding and shear band evolution. Furthermore， parametric analysis confirmed that introducing a "negative dilatancy angle" to simulate shear contraction in loose colluvium is essential for reflecting excess pore water pressure generation and subsequent fluidization. For 3D simulation， pre- and post-failure elevation analyses were integrated with drilling data to construct a geometric model of the Tuchang landslide. The severe rainfall was directly converted into an initial high groundwater level for 3D fluid-solid coupling analysis. Results indicate that， driven by this high water level， shear contraction triggers local excess pore water pressure， reducing effective stress and leading to a large-scale debris-flow-like movement. Finally， a confusion matrix assessment showed the baseline model achieved an Intersection over Union (IoU) of 46.11%， with 58.36% precision and 68.71% recall. This demonstrates a reasonable spatial correspondence with the actual post-failure deposition， providing a specific reference for future landslide mitigation.
