[A-HW17] Near Surface Investigation and Modeling for Groundwater Resources Assessment and Conservation

[AHW17-P09] Estimating Hydraulic Heterogeneity Using DenseNet and Hydraulic Tomography

*Yukai Chen¹, Jui-Pin Tsai¹ (1.National Taiwan University)

Keywords:Hydraulic tomography, Hydrogeological heterogeneity, Convolutional neural network, Deep learning

Understanding subsurface geological structures is essential for groundwater resource development. Hydraulic tomography (HT) is a mature technique for delineating three-dimensional hydrogeological parameter fields. Successive linear estimator (SLE) is the commonly utilized inverse method for HT, a geostatistical approach that has been extensively validated and applied across diverse scale scenarios. However, as the number of parameters and observation data increases, SLE's computational efficiency decreases. To address this issue, we developed a convolutional neural network based on HT (i.e., HT-NN) to replace the SLE algorithm for converting head/drawdown data to hydraulic heterogeneity. To develop HT-NN, a groundwater flow simulator, VSAFT2, was used to generate extensive data pairs, including head data, hydraulic conductivity (K) fields, and specific storage (Ss) fields. These data pairs were subsequently utilized as training datasets for training HT-NN. The study results demonstrate that the HT-NN successfully converts the head dataset to K/Ss fields, and the parameter fields estimated by the HT-NN capture the spatial characteristics of the original parameter fields. The results further highlight the potential of HT-NN as an alternative to the SLE algorithm, offering substantial advantages for assessing hydraulic heterogeneity and addressing related groundwater research concerns.