Large-scale machine-learning molecular dynamics simulation of primary radiation damage in tungsten

Simulating collision cascades and radiation damage poses a long-standing challenge for existing interatomic potentials, both in terms of accuracy and efficiency. Machine-learning-based interatomic potentials have shown sufficiently high accuracy for radiation damage simulations, but most existing ones are still not efficient enough to model high-energy collision cascades with sufficiently large space and timescales. To this end, we here extend the highly efficient neuroevolution potential (NEP) framework by combining it with the Ziegler-BiersackLittmark (ZBL) screened nuclear repulsion potential, obtaining a NEP-ZBL framework. We train a NEP-ZBL model for tungsten and demonstrate its accuracy in terms of the elastic properties, melting point, and various energetics of defects that are relevant for radiation damage. We then perform large-scale molecular dynamics simulations with the NEP-ZBL model with up to 8.1 million atoms and 240 ps (using a single 40-GB A100 GPU) to study the difference of primary radiation damage in both bulk and thin-foil tungsten. While our findings for bulk tungsten are consistent with existing results simulated by embedded atom method models, the radiation damage differs significantly in foils and shows that larger and more vacancy clusters as well as smaller and fewer interstitial clusters are produced due to the presence of a free surface.