Data for: Hydrodynamic-guided graph learning bridges data-driven and physics-informed artificial intelligence

Description 1. Physics-based Model (HydroGeoSphere) A. Dataset Description: The dataset in this repository mainly consists of the numerical simulations described in the main paper. All the data used for figures are stored in this repository, and the file names correspond to the figure names in the main paper. For the simulation files, the grok (.grok) file in the dataset are the pre-processor. To perform the simulations and visualize their results, three executables (GROK.EXE, HGS.EXE, HSPLOT.EXE) are required. Any additional details can be obtained from the corresponding author (Hyoun-Tae Hwang). B. Modelling Software Version Requirements: HGS version 2023 or higher is required for running the forward simulations. C. License Notice: Please note that permission is required for redistributing the data, and the material should not be used for commercial purposes. Additionally, HGS is a commercial software, and for HGS license purchase, please contact www.aquanty.com. 2. Deep Learning Model (Graph Attention Network) A. Code Description: This repository contains the data preprocessing and deep learning model training codes used in the manuscript entitled "Hydrodynamic-guided deep learning for three-dimensional spatiotemporal prediction of subsurface contamination", currently under review in Nature Water. The proposed framework converts HydroGeoSphere (HGS) simulation outputs into graph-structured datasets. A hydrodynamic-guided Graph Attention Network (GAT) is then employed to predict the spatiotemporal evolution of subsurface contamination. The repository includes both the preprocessing workflow for graph dataset construction and the training framework for the hydrodynamic-guided GAT model. Together, these codes reproduce the complete data processing and modeling pipeline presented in the manuscript. The repository is organized into two major components: 1) Data preprocessing and 2) GAT model training B. Notes: The final GAT model is trained using the main contamination area. Accordingly, graph-structured data, including target contaminant concentrations, hydrodynamic distances, and sink-related features, are constructed within this domain. However, hydrodynamic distance calculation requires the original domain representing the entire Wonju industrial complex. Groundwater flow is not restricted to the main contamination area and may leave and later re-enter the domain. Therefore, preprocessing Steps 3–5 generate additional velocity, graph, and target datasets for the origin domain. These datasets are not used directly for model training but serve as inputs for hydrodynamic distance calculation in Step 6. Hydrodynamic distances are first computed over the origin domain and then indexed to the main contamination area. The resulting distance datasets are subsequently used as model inputs. Thus, preprocessing outputs labeled origin are intermediate products for hydrodynamic distance calculation rather than extensions of the final training domain.