Physics-guided deep learning surrogate model with graph attention for long-term radionuclide transport prediction in deep geological repositories

Abstract

To ensure millennial isolation of high-level waste, Deep geological repositories (DGR) safety assessments depend on process-based simulators such as Parallel flow and reactive transport model (PFLOTRAN). Yet, without feasible long-term monitoring, the high computational burden of these models becomes a bottleneck for iterative scenario testing and policy decisions. To overcome this, we developed a graph attention-based physics-guided deep learning (GAT-PGDL) surrogate that embeds decay-diffusion-sorption equations. U-238 and Th-230 transport was simulated for 5000 years in DGRs, with release commencing at year 2000 across ten monitoring nodes. On a single-node workstation, PFLOTRAN requires similar to 5760 min per scenario, whereas the GAT-PGDL trains once (similar to 94 min) and infers in seconds, delivering similar to 61 x per-scenario speedup. To assess reliability and interpretability, we employed split conformal prediction for uncertainty quantification, perturbation-based sensitivity analysis, and feature importance analysis. The GAT-PGDL produced 95 % prediction intervals and highlighted sorption and bulk density as key transport controls. For generalization, performance was compared with a data-driven surrogate under scenarios with altered material properties and earlier release times, where the GAT-PGDL outperformed the data-driven model, maintaining R-2 and NSE > 0.98. These results establish GAT-PGDL as a fast, accurate, and physically reliable surrogate for long-term DGR safety assessments.