An Accurate and Transferable Machine-Learning Interatomic Potential for Silicon

Abstract: The development of modern ab initio methods has rapidly increased our understanding of physics, chemistry and materials science. Unfortunately, intensive ab initio calculations are intractable for large and complex systems. On the other hand, empirical force fields are less accurate with poor transferability even though they are efficient to handle large and complex systems. The recent development of machine-learning based neural-network (NN) for local atomic environment representation of density functional theory (DFT) has offered a promising solution to this long-standing challenge. Si is one of the most important elements in science and technology, however, an accurate and transferable interatomic potential for Si is still lacking. Here, we develop a generalized NN potential for Si, which correctly predicts the Si(111)-(7x7) ground-state surface reconstruction for the first time and accurately reproduces the DFT results in a wide range of complex Si structures. We envision similar developments will be made for a wide range of materials systems in the near future.