Insights into the primary radiation damage of silicon by a machine learning interatomic potential

We develop a silicon Gaussian approximation machine learning potential suitable for radiation effects, and use it for the first <i>ab initio</i> simulation of primary damage and evolution of collision cascades. The model reliability is confirmed by good reproduction of experimentally measured threshold displacement energies and sputtering yields. We find that clustering and recrystallization of radiation-induced defects, propagation pattern of cascades, and coordination defects in the heat spike phase show striking differences to the widely used analytical potentials. The results reveal that small defect clusters are predominant and show new defect structures such as a vacancy surrounded by three interstitials. Quantum-mechanical level of accuracy in simulation of primary damage was achieved by a silicon machine learning potential. The results show quantitative and qualitative differences from the damage predicted by any previous models.