Machine learning force field for thermal oxidation of silicon.

Looking back at seven decades of highly extensive application in the semiconductor industry, silicon and its native oxide SiO2 are still at the heart of several technological developments. Recently, the fabrication of ultra-thin oxide layers has become essential for keeping up with trends in the down-scaling of nanoelectronic devices and for the realization of novel device technologies. With this comes a need for better understanding of the atomic configuration at the Si/SiO2 interface. Classical force fields offer flexible application and relatively low computational costs, however, suffer from limited accuracy. Ab initio methods give much better results but are extremely costly. Machine learning force fields (MLFF) offer the possibility to combine the benefits of both worlds. We train a MLFF for the simulation of the dry thermal oxidation process of a Si substrate. The training data are generated by density functional theory calculations. The obtained structures are in line with ab initio simulations and with experimental observations. Compared to a classical force field, the most recent reactive force field, the resulting configurations are vastly improved. Our potential is publicly available in an open-access repository.