Construction and simulation proofs of reliable high-dimensional neural network atomic potentials*

High-dimensional neural network atomic potentials (HDNNP) [1] has attracted attention because of their potential to achieve reliability and computational efficiency simultaneously. In this talk, we show our attempts to apply the HDNNP to several different topics, and discuss how to construct reliable HDNNP and verify the reliability through simulation proofs.
We show our results on Li ion diffusion in amorphous Li₃PO₄ [2] and thermal conductivities of wurtzite GaN and silicon crystals [3]. In both cases, we achieved good agreement with calculations within the density functional theory. Then we discuss the atomic energy mapping inferred by HDNNP [4]. We show that the energy mapping can be improved by choosing the training set carefully and monitoring the atomic energy during the training procedure. In addition, we will also touch on the HDNNP to examine the atomic structures of a complex four-element system, Au/Li₃PO₄ interfaces [5].
[1] J. Behler and M. Parrinello, Phys. Rev. Lett. 98, 146401 (2007).
[2] W. Li, et al., J. Chem. Phys. 147, 214106 (2017).
[3] E. Minamitani, et al., Appl. Phys. Express 12, 095001 (2019).
[4] D. Yoo, et al., Phys. Rev. Mater. 3, 093802 (2019).
[5] K. Shimizu, et al., in preparation.