講演情報
[20a-P04-2]High-performance quadratic spectral neighbor analysis potential for Al-H binary system to investigate hydrogen embrittlement
〇(M1)Yuxi Liu1, Koji Shimizu1, Satoshi Watanabe1 (1.UTokyo)
キーワード：
machine-learning potentials、hydrogen embrittlement、atomistic simulations
Aluminum and its high-strength alloys suffer from hydrogen embrittlement (HE), which is extensively studied but its mechanisms have not been comprehensively understood yet. Recently, machine learning potentials (MLPs) combined with molecular dynamics (MD) simulations provide us a feasible way to study HE through MD simulations with larger spatial and time scales and high accuracy. Here, we constructed a MLP for Al-H binary system within the framework of quadratic spectral neighbor analysis potential (qSNAP) to investigate HE in Al. The root-mean-square errors (RMSE) of total energies and atomic forces of our qSNAP model, compared with DFT results, were 2.4 meV/atom and 58 meV/ang, respectively. The qSNAP model agreed well with those of DFT on various materials properties. In addition, a decrease of hydrogen migration barrier between tetrahedral and octahedral sites from 0.164 eV to 0.110 eV is observed with the increase of temperature from 100K to 900K.