Author(s): Min Pang; Erhu Du; Chunmiao Zheng

Linked Author(s):

Keywords: Groundwater contamination; Contaminant transport modeling; Source attribution; Deep learning; Graph neural network

Abstract: Contaminant transport modeling in heterogeneous aquifers from multiple sources presents a complex and challenging problem in groundwater management. Traditional modeling approaches often encounter difficulties in capturing the intricacies of contaminant dispersion, limitations in available data, and the substantial computational demands. In this work, we propose a new deep learning method, known as the attention-based graph neural network (aGNN), to model contaminant transport with limited monitoring data. In four case studies involving varying monitoring networks in heterogeneous aquifers, aGNN is shown to outperform LSTM (long short-term memory) and CNN (convolutional neural network) based methods in multistep predictions. Furthermore, explanatory analysis based on aGNN quantifies the influence of each contaminant source, aligning with a physics-based model and achieving high R2 values. The key advantage of aGNN is its ability to achieve high predictive accuracy across multiple scenarios while significantly reducing computational demands. Overall, our findings demonstrate that aGNN is an efficient and robust tool for nonlinear spatiotemporal learning in contaminant transport within the subsurface, offering promising applications in groundwater management.

DOI: https://doi.org/10.3850/iahr-hic2483430201-315

Year: 2024