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.

Interpretable Machine Learning to Discover Perovskites with High Spontaneous Polarization

Machine Learning Design of Perovskite Catalytic Properties

Machine learning identified materials descriptors for ferroelectricity

Highly versatile and accurate machine learning methods for predicting perovskite properties

Identification of novel organic ferroelectrics: A study combining importance sampling with machine learning

A Domain Knowledge Enhanced Machine Learning Method to Predict the Properties of Halide Double Perovskite A2B+B3+X6

Accelerated Discovery of Stable Lead-Free Hybrid Organic-Inorganic Perovskites Via Machine Learning

Machine learning facilitated by microscopic features for discovery of novel magnetic double perovskites

Rationalizing Perovskite Data for Machine Learning and Materials Design

Topological feature engineering for machine learning based halide perovskite materials design

Machine learning for imbalanced datasets: Application in prediction of 3d-5d double perovskite structures

Accelerated design and discovery of perovskites with high conductivity for energy applications through machine learning

Machine learning for halide perovskite materials

Interpretable machine learning model of effective mass in perovskite oxides with cross-scale features

Voting Data-Driven Regression Learning for Discovery of Functional Materials and Applications to Two-Dimensional Ferroelectric Materials

Photoelectrochemical Properties, Machine Learning, and Symbolic Regression for Molecularly Engineered Halide Perovskite Materials in Water

Applications of machine learning in perovskite materials

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

Designing New Improper Ferroelectrics with a General Strategy

Machine learning stability and band gap of lead-free halide double perovskite materials for perovskite solar cells

Machine Learning for Perovskite Solar Cells and Component Materials: Key Technologies and Prospects