Designing Two-Dimensional Halide Perovskites Based on High-Throughput Calculations and Machine Learning

Abstract

The interactions between ions and the low-dimensional halide perovskites are critical to realizing the next-generation energy storage devices such as photorechargeable ion batteries and ion capacitors. In this study, we performed high-throughput calculations and machine-learning analysis for ion adsorption on two-dimensional A2BX4 halide perovskites. The first-principles calculations obtained an initial data set containing adsorption energies of 640 compositionally engineered ion/perovskite systems with diverse ions including Li+, Zn2+, K+, Na+, Al3+, Ca2+, Mg2+, and F–. The machine learning algorithms including k-nearest neighbors (KNN), Kriging, Random Forest, Rpart, SVM, and Xgboost algorithms were compared, and the Xgboost algorithm achieved the best accuracy (r = 0.97, R 2 = 0.93) and was selected to predict the virtual design space consisting of 11 976 ion/perovskite systems. The features were then analyzed and ranked according to their Pearson correlations to the output values. In particular, to better understand the features, diverse feature selection methods were employed to comprehensively evaluate the features. The machine-learning-predicted virtual design space was subsequently screened to select stable lead-free ion/perovskite systems with suitable band gaps and halogen mixing features. The present study provides a theoretical foundation to design halide perovskite materials for ion-based energy storage applications such as secondary ion batteries, ion capacitors, and solar-rechargeable batteries.

Small data machine learning in materials science

“…High-throughput computations and experiments have become significant methods to provide sufficient materials data for machine learning research. Hu et al 76 obtained 640 2D halide perovskites A 2 BX 4 (A = Li, Na, K, Rb, Cs; B = Ge, Sn, Pb; X = F, Cl, Br, I) and corresponding adsorption energies with Li + , Zn 2+ , K + , Na + , Al 3+ , Ca 2+ , Mg 2+ , and F − by using high-throughput computations. After filtering out 13 descriptors with the Pearson correlation coefficient, k-nearest neighbors (KNN), Kriging, Random Forest, Rpart, SVM, and XGBoost were adopted for modeling.…”

Machine learning in energy chemistry: introduction, challenges and perspectives

“…142 Using a combination of DFT optimization and machine learning prediction, they determined the range of tolerance factors, octahedral factors, metal electronegativity, and polarizability of potentially promising HOIP organic molecules and selected 3 thermal and environmental stable lead-free HOIPs with appropriate bandgaps from 5158 candidates. In addition, there have been research efforts that combine machine learning and DFT, 143 for discovering lead-free hybrid perovskite, 144 two-dimensional lead-free perovskite 145 and others. 118,146 These works provide a solid foundation for discovering more efficient and stable lead-free halide perovskites.…”

Accelerating the Discovery of Transition Metal Borides by Machine Learning on Small Data Sets

“…It is obtained through convex hull (CH) analysis, 29 but this method is complicated and costly due to the necessity of manual programming and complex data processing. Thanks to the advanced ML algorithms based on big data sets, 30 novel material discovery and design have been accelerated considerably, for example, the predictions of perovskites, 31 alloys, 32 polymers, 33 have made great contributions in many fields, such as electronic devices, 34 energy storage, 35 and electrocatalysts. 36 However, the requirement of large data volume and the unreliability of the data sets 31 hinder the application of ML methods for accurately predicting material properties on limited data set (set size < 10 3 ) material systems like MABs, MXenes, MBenes, etc.…”

“…Thanks to the advanced ML algorithms based on big data sets, 30 novel material discovery and design have been accelerated considerably, for example, the predictions of perovskites, 31 alloys, 32 polymers, 33 have made great contributions in many fields, such as electronic devices, 34 energy storage, 35 and electrocatalysts. 36 However, the requirement of large data volume and the unreliability of the data sets 31 hinder the application of ML methods for accurately predicting material properties on limited data set (set size < 10 3 ) material systems like MABs, MXenes, MBenes, etc. Fortunately, accurate prediction of material properties, including formation energy, 37 band gap, 38 bulk modulus, 39 and vibrational entropy, 40 now can be achieved by using the small-data set ML methods.…”

“…It is obtained through convex hull (CH) analysis, but this method is complicated and costly due to the necessity of manual programming and complex data processing. Thanks to the advanced ML algorithms based on big data sets, novel material discovery and design have been accelerated considerably, for example, the predictions of perovskites, alloys, polymers, etc. Besides, ML algorithms have made great contributions in many fields, such as electronic devices, energy storage, and electrocatalysts .…”