Accelerating Pmn21-BAlNP properties prediction by machine learning based on first-principles calculation. (March 2019)

• Record Type: Journal Article
• Authors: Zhu, Chuanshuai; Yang, Ruike; chai, Bao; Wei, Qun; Zhang, Dongyun
• Is Part Of: Journal of physics and chemistry of solids. Volume 126(2019)
• Page Start: 224
• Page End: 233
• Publication Date: 2019-03
• Subjects: Pmn21-BxAl1-xNyP1-y -- Mechanical properties -- Optoelectronic properties -- First-principles -- Machine learning; Solids -- Periodicals; Solides -- Périodiques; Solids; Periodicals; 530.41
• DOI: 10.1016/j.jpcs.2018.11.024
• Languages: English
• ISSNs: 0022-3697

Abstract

Abstract: In this paper, the CASTEP program and machine learning algorithm are used to study Pmn 21 -B x Al1- x N y P1- y . The mechanical stability, thermodynamic stability and molecular dynamics stability of several typical structures of Pmn 21 -B x Al1- x N y P1- y are estimated by CASTEP program. The results show that all the structures are stable. The data such as lattice constants and elastic properties of Pmn 21 -B x Al1- x N y P1- y are used as training sets of machine learning, then, the average absolute percent error (MAPE) of each model is calculated by using 7-fold cross-validation, and the lattice constant prediction model and shear modulus prediction model of Pmn 21 -B x Al1- x N y P1- y are selected respectively. Linear regression algorithm is used for modeling lattice constants, and Xgboost algorithm is used for modeling shear modulus. On the test sets, it is found that the error is within the acceptable range by comparing the data predicted by the models with the real data. Therefore, the machine learning prediction models of lattice constants, and shear modulus G have strong generalization ability. The electrical properties of Pmn 21 -B x Al1- x N y P1- y are calculated by using hybrid PBE0 functional. The direct band gaps are found in AlN, BP, AlP, B0.25 Al0.75 N, B0.75 Al0.25 P, B0.5 Al0.5 P, B0.25 Al0.75 P, AlN0.75 P0.25, AlN0.5 P0.5 and AlN0.25 P0.75 . Highlights: Pmn 21 -B x Al1- x N y P1- y are not researched by others. Machine learning prediction models of lattice constants, and shear modulus G have strong generalization ability. Xgboost algorithm is used for modeling shear modulus.