Enhancing hydraulic conductivity field characterization through integration of hydraulic head and tracer data using multi-modal neural network models

Abstract

Characterizing heterogeneous conductivity field is essential for effective groundwater management and controlling contaminant events. The characterization method is currently advancing towards two promising directions: (1) integration of various types of data, such as hydraulic head (H) and tracer concentration (C) data; (2) usage of machine learning methods of AI area. However, no machine learning model has been proposed to integrate H and C data for aquifer characterization effectively. The two data types have different forms: H data being spatial and C data being temporal. This discrepancy creates challenges for effective integration.

We developed three machine learning models—HydroCNN, HC-Net1, and HC-Net2. The HydroCNN model could effectively predict hydraulic conductivity (K) fields from H data alone and thus is used as the baseline for evaluating the following models. HC-Net1 and HC-Net2 models are multi-modal neural network models with different architectures. These multi-modal architectures incorporate both convolutional neural network and fully connected neural network modules, designed to integrate H and C data to enhance characterization accuracy. Results indicate that HC-Net2 significantly outperforms the other models, highlighting its capability to leverage the strengths of both data types effectively. Notably, the HC-Net2 model’s improvement is most significant in scenarios where models relying solely on H data perform poorly.