Data-Driven Prediction of Tritium Transport Parameters for Fusion Reactor Fuel Cycle Modeling

Abstract

Accurate tritium transport predictions are critical for fuel cycle optimization, tritium inventory control, and safety assessment in magnetic confinement fusion devices such as ITER and future DEMO reactors. However, inverse identification of diffusion and trapping parameters from experimental release measurements is often ill-posed due to sparse data, limited temperature coverage, or uncertain trapping characteristics. Here, we develop a data-driven inverse modeling framework based on deep neural networks (DNNs) to infer key parameters governing one-dimensional tritium transport in solids described by the McNabb-Foster diffusion-trapping model. Physically consistent synthetic datasets are generated by numerically solving coupled diffusion-trapping equations over a broad parameter space. Cumulative tritium release at multiple temperatures serves as the network input. Two prediction scenarios are examined: (i) known trapping parameters and (ii) unknown trapping parameters. In both cases, the DNNs accurately predict the diffusion pre-factor and activation energy, achieving average relative errors below 3.5%. Model interpretability is assessed using SHAP (SHapley Additive exPlanations) analysis, indicating that inversion is predominantly governed by high-temperature release behavior. The resulting transport parameters can provide physically informed inputs for fusion reactor fuel cycle simulations, enabling improved estimates of tritium retention, recycling, and plasma fueling efficiency. This work demonstrates a robust and efficient methodology for integrating experimentally accessible release data into reactor-scale tritium inventory assessments.