Data-driven classification and prediction of all-inorganic Cs-Pb-Br perovskite crystal structures

All-inorganic perovskites based on cesium-lead-bromine (Cs-Pb-Br) have been a prominent research focus in optoelectronics in recent years. The optimisation and tunability of their macroscopic properties exploit the conformational flexibility, resulting in various crystal structures. Varying synthesis parameters can yield distinct crystal structures from Cs, Pb, and Br precursors, and manually exploring the relationship between these synthesis parameters and the resulting crystal structure is both labour-intensive and time-consuming. Machine learning (ML) can rapidly uncover insights and drive discoveries in chemical synthesis with the support of data, significantly reducing both the cost and development cycle of materials. Here, we gathered synthesis parameters from published literature (220 synthesis runs) and implemented eight distinct ML models, including eXtreme Gradient Boosting (XGB), Decision Tree (DT), Support Vector Machine (SVM), Random Forest (RF), Naïve Bayes (NB), Logistic Regression (LR), Gradient Boosting (GB), and K-Nearest (KN) to classify and predict Cs-Pb-Br crystal structures from given synthesis parameters. Validation accuracy, precision, F1 score, recall, and average area under the curve (AUC) are employed to evaluate these ML models. The XGB model exhibited the best performance, achieving a validation accuracy of 0.841. The trained XGB model was subsequently utilised to predict the structure from 10 experimental runs using a randomised set of parameters, achieving a testing accuracy of 0.8. The results indicate that the Cs/Pb molar ratio, reaction time, and the concentration of organic compounds (ligands) play crucial roles in synthesising various crystal structures of Cs-Pb-Br. This study demonstrates a significant decrease in effort required for experimental procedures and builds a foundational basis for predicting crystal structures from synthesis parameters.