Primary radiation damage in silicon from the viewpoint of a machine learning interatomic potential

Characterization of the primary damage is the starting point in describing and predicting the irradiation-induced damage in materials. So far, primary damage has been described by traditional interatomic potentials in molecular dynamics simulations. Here, we employ a Gaussian approximation machine-learning potential (GAP) to study the primary damage in silicon with close to ab initio precision level. We report detailed analysis of cascade simulations derived from our modified Si GAP, which has already shown its reliability for simulating radiation damage in silicon. Major differences in the picture of primary damage predicted by machine-learning potential compared to classical potentials are atomic mixing, defect state at the heat spike phase, defect clustering, and recrystallization rate. Atomic mixing is higher in the GAP description by a factor of two. GAP shows considerably higher number of coordination defects at the heat spike phase and the number of displaced atoms is noticeably greater in GAP. Surviving defects are dominantly isolated defects and small clusters, rather than large clusters, in GAP's prediction. The pattern by which the cascades are evolving is also different in GAP, having more expanded form compared to the locally compact form with classical potentials. Moreover, recovery of the generated defects at the heat spike phase take places with higher efficiency in GAP. We also provide the attributes of the new defect cluster that we had introduced in our previous study. A cluster of four defects, in which a central vacancy is surrounded by three split interstitials, where the surrounding atoms are all 4-folded bonded. The cluster shows higher occurrence in simulations with the GAP potential. The formation energy of the defect is 5.57 eV and it remains stable up to 700 K, at least for 30 ps. The Arrhenius equation predicts the lifetime of the cluster to be 0.0725 mu s at room temperature.