Understanding hydrogen trapping in tungsten is crucial for accurate modeling of plasma-wall interactions in fusion devices. In this study, we developed a deep learning model to predict hydrogen trapping sites, which serve as essential input parameters for kinetic Monte Carlo (kMC) simulations. The model employs a U-Net–based convolutional neural network that directly maps three-dimensional potential energy distributions to trapping site positions. Training data were generated from atomistic calculations using the embedded atom method (EAM) potential, and ground-truth trapping sites were systematically identified by force relaxation. The trained model achieved an F1 score of approximately 0.76, with most predicted sites coinciding with the true minima within ± 1 voxels (1 voxel = 0.1 Å per side). Visual comparisons confirmed the ability of the model to capture both global and local features of the potential energy landscape. In terms of efficiency, the proposed approach reduced prediction time by more than three orders of magnitude compared with conventional force-relaxation searches, enabling predictions in less than one second on a GPU. These results demonstrate that deep learning provides an accurate and computationally efficient method for identifying trapping sites. Future extensions include incorporating more complex defect structures and integrating the model into molecular dynamics-kMC hybrid frameworks for large-scale plasma-wall interaction simulations.
Saito et al. (Thu,) studied this question.