Data-efficient machine-learning interatomic potential for studying radiation effects in germanium

Abstract Exposure to harsh radiation environments leads to displacement damage in semiconductor materials, ultimately degrading the device performance. Molecular dynamics (MD) is a powerful method for simulating the dynamic processes of radiation-induced defect generation, clustering, and evolution, which are often beyond the reach of experimental observation. In this work, we introduce a machine-learning Gaussian approximation potential for germanium, specifically designed to describe properties related to radiation-induced collision cascades and the resulting damage. A repulsive potential is incorporated to accurately capture short-range interactions during cascade simulations. In addition, we reproduce elastic, thermal, and vibrational properties, and accurately capture the energetics of vacancies and self-interstitials. These high-fidelity predictions are comparable to ab initio calculations while requiring only a compact training database. This potential is expected to facilitate accurate MD simulations of radiation damage in germanium with greatly improved accuracy compared to empirical potentials.