[A-HW25] Near Surface Investigation and Modeling for Groundwater Resources Assessment and Conservation

[AHW25-P02] Hydraulic Heterogeneity Estimation using Hydraulic Tomography Neural Network (HT-NN): A Field Study in Taiwan.

*Zi-Yan Chou¹、Jui-Pin Tsai¹ (1.National Taiwan University)

キーワード：Hydraulic Heterogeneity, Hydraulic Tomography, Successive Linear Estimation, Convolutional Neural Network

Characterizing the heterogeneity of hydraulic parameters (K or Ss) in the subsurface is crucial for contamination remediation. The Hydraulic Tomography (HT) is a well-developed approach for estimating three-dimensional hydrogeological parameter fields by capturing variations of groundwater head stimulated by pumping/injection events. The core of the HT is the successive linear estimation (SLE), a well-proven geostatistical inverse method in various scale problems. However, the computational efficiency decreases as the number of grid or head observations increases. In order to increase the efficiency of estimation, we propose a HT-based convolutional neural network (HT-NN) to replace SLE for converting groundwater head variations into hydraulic heterogeneity. In this study, we conducted a HT field test at a contamination field in Miaoli, Taiwan. We generated a large number of random fields based on core sample data near the site. For each field, we executed forward simulation to obtain head variation under the same injection configuration used in the field. These synthetic datasets, consisting of random hydraulic fields (outputs) and corresponding head variations (inputs), were used to train the HT-NN model. We then compared HT-NN predictions with traditional SLE results using filed observations. The result show that HT-NN successfully converts the groundwater head variations into hydraulic parameters fields, proving that the development of HT-NN can be an efficient tool to depict the subsurface heterogeneity in real-time, and benefiting the strategies of remediation process over time.