Integrative approach of machine learning and symbolic regression for stability prediction of multicomponent perovskite oxides and high-throughput screening

Created by W.Langdon from gp-bibliography.bib Revision:1.8129

@Article{ZHANG:2024:commatsci,

• author = "Zhaosheng Zhang and Yingjie Zhang and Sijia Liu",
• title = "Integrative approach of machine learning and symbolic regression for stability prediction of multicomponent perovskite oxides and high-throughput screening",
• journal = "Computational Materials Science",
• year = "2024",
• volume = "236",
• pages = "112889",
• month = mar,
• keywords = "genetic algorithms, genetic programming, Symbolic regression, Machine learning, High-throughput, Perovskite oxides",
• ISSN = "0927-0256",
• URL = "https://www.sciencedirect.com/science/article/pii/S0927025624001101",
• DOI = "doi:10.1016/j.commatsci.2024.112889",
• abstract = "This work unfolds a robust and interpretable strategy for evaluating the stability and potential photovoltaic application of 6526 multicomponent perovskite oxides, employing a synergetic methodology that intertwines advanced machine learning (ML) algorithms and symbolic regression based on genetic programming. Initially, ML algorithms, namely XGBoost, LightGBM, and random forest, were harnessed, with elemental oxidation state and electronegativity serving as input features, achieving R2 values of 0.98, 0.98, and 0.74, respectively, on the test set for predicting the formation enthalpy, a criterion for perovskite stability. Despite the amplified interpretability offered by SHAP analysis, the inherent black-box nature of ML obfuscates a transparent understanding of intrinsic relations between input features and performance. To surmount this, symbolic regression introduced not only elucidates a clear functional relationship between input features and perovskite stability but also achieves a commendable R2 of 0.79 on the test set. Subsequent high-throughput screening, based on perovskite stability ranking, designated the top 500 stable perovskites for band gap calculation using the PBE functional, wherein DyNdHf2O6, CeEuAl2O6, and CeSrAl2O6 emerged as potential candidates for photovoltaic applications and were subjected to further electronic structure simulations employing the HSE06 functional, encompassing density of states, band structure, charge density, and optical absorption spectra. Ultimately, CeEuAl2O6, boasting an optical direct bandgap of 2.31 eV and minimal electron-hole wavefunction overlap, stands out as the prime choice for photovoltaic materials. This research not only pioneers the exploration of enhancing the interpretability of ML but also propels theoretical guidance for the evolution of photovoltaic cells by bridging predictive modeling with high-throughput screening",

}

Genetic Programming entries for Zhaosheng Zhang Yingjie Zhang Sijia Liu

Citations