跳到主要內容 :::

年度	113	專案性質	實驗性質
專案類別	研究專案	研究主題	整治
申請機構	國立台灣大學	申請系所	生物環境系統工程學系
專案主持人	蔡瑞彬	職等/職稱	副教授
專案中文名稱	基於深度學習之三維水力掃描技術發展與現地測試
中文關鍵字	深度學習，水力掃描，三維異質水文地質參數場
專案英文名稱	Development and Field Testing of a Deep Learning Approach for 3D Hydraulic Tomography
英文關鍵字	Deep learning, hydraulic tomography, 3D heterogeneous hydrogeological parameter fields
執行金額
執行期間	2025/1/1 至 2025/12/31
計畫中文摘要	本研究聚焦於污染場址整治過程中的一項關鍵挑戰：如何精確掌握現地水文流場，以提升藥劑與污染團流佈的預測能力。傳統方法多假設水文地質場為均質條件，當面對高度異質性的土層時，常導致預測結果與實際情況出現明顯偏差，進而影響整治決策的效能與效率。為突破此限制，本研究導入水力掃描技術（Hydraulic Tomography），作為推估異質性三維水文地質參數場的核心方法。該技術透過多點循序抽水／注水與水位觀測，可反演三維水力傳導係數與比儲水係數場，其有效性已在多尺度應用中獲得驗證。然而，水力掃描法在實務上需處理大量時序觀測資料與高維參數空間，導致計算成本高昂，限制其於整治場域中的即時應用。為解決此問題，本研究提出結合深度學習與水力掃描的「水力掃描神經網路（Hydraulic Tomography Neural Network, HT-NN）」技術，作為高效率反演異質性參數場的創新方法。針對地層異質性的複雜性，本研究設計高效的訓練資料生成策略，涵蓋多樣化地質條件，並強化模型對局部水文結構的辨識能力，使深度學習模型能更真實地反映現地場的異質性。HT-NN 結合水力層析理論、深度學習模型與條件模擬方法，可快速重建異質地層中水力傳導係數（K）的三維時序分布。本研究於臺大農場入滲儀建置由紅土與川砂交互堆疊之異質滲透場，進行多次水力掃描試驗，以分析藥劑灌注前後K場變化。結果顯示，灌注藥劑後孔隙阻塞使鄰近灌注井區域之 K 值顯著下降。而HT-NN 在訓練與測試資料的 R² 達到 0.879，顯示其能有效學習水位與 K 場之非線性關係；應用於藥劑灌注後入滲儀之水力掃描資料分析時，與循序線性推估器（SLE）所得 K 場之比較結果 R² = 0.626，證實模型能重現孔隙阻塞造成的異質結構與滲透變化。由此結果可知，條件模擬方法的引入，使 HT-NN 在訓練資料生成階段得以保留地質統計特性，並提升模型於低透水區及高-低 K 值過渡帶的推估穩定性。整體結果顯示，HT-NN 同時具備高效率與高解析度，為污染整治過程中三維水力場動態監測與決策支援的重要分析工具。
計畫英文摘要	This study focuses on a critical challenge in contaminated-site remediation: accurately characterizing in-situ hydrological flow fields to improve the prediction of reagent transport and contaminant plume migration. Conventional approaches often assume homogeneous hydrogeological conditions; however, when dealing with highly heterogeneous formations, such assumptions frequently lead to significant discrepancies between predicted and observed results, thereby reducing the effectiveness and efficiency of remediation decisions. To overcome this limitation, this study adopts Hydraulic Tomography (HT) as the core technique for estimating three-dimensional heterogeneous hydrogeological parameter fields. HT employs sequential multi-well pumping/injection and head observations to reconstruct 3D hydraulic conductivity (K) and specific storage (Ss) fields, and its effectiveness has been validated across multiple spatial scales. Nonetheless, the practical implementation of HT requires processing large time-series datasets and solving high-dimensional parameter spaces, resulting in high computational costs that hinder its real-time application in remediation sites. To address this issue, this study proposes an innovative technique, the Hydraulic Tomography Neural Network (HT-NN), which integrates deep learning with hydraulic tomography for efficient inversion of heterogeneous parameter fields. To handle the complexity of subsurface heterogeneity, a high-efficiency training data generation strategy was developed to encompass diverse geological conditions and enhance the model’s capability to recognize local hydraulic structures, allowing the deep-learning framework to better capture the realistic spatial variability of in-situ systems. The HT-NN integrates hydraulic-tomography theory, deep-learning models, and conditional simulation, enabling rapid reconstruction of time-dependent three-dimensional hydraulic conductivity (K) fields in heterogeneous media. A layered infiltration tank composed of alternating lateritic clay and river sand was constructed at National Taiwan University to simulate a realistic heterogeneous permeable system. Multiple hydraulic tomography tests were conducted to analyze K-field variations before and after reagent injection. The results showed that pore clogging following reagent injection caused a significant decrease in K values near the injection zone. The HT-NN achieved an R² of 0.879 for training and testing datasets, demonstrating its strong capability in learning the nonlinear relationship between hydraulic heads and K fields. When applied to post-injection infiltration data, the comparison between HT-NN and the Successive Linear Estimator (SLE) yielded an R² of 0.626, confirming that the model successfully reproduced the heterogeneous structures and permeability variations induced by pore clogging. These results indicate that the incorporation of conditional simulation enables HT-NN to preserve geostatistical characteristics during training-data generation and enhances model stability in low-permeability regions and transition zones between high and low K values. Overall, the HT-NN demonstrates both high efficiency and high resolution, serving as a powerful analytical tool for dynamic three-dimensional hydraulic-field monitoring and decision support in contaminated-site remediation.