A2BB’X6/ABX3-type High-performance perovskites screening based on ensemble learning and high throughput screening

Abstract

In this work, we have built a complete set of the ensemble learning model through data cleaning, feature engineering, model training, and evaluation. The model has predicted the formation energy (E_form), energy above hull (E_hull), and band gap of perovskites with an accuracy close to that of the first-principles calculation of a Mean Absolute Error of 0.146 eV, 0.041 eV, and 0.396 eV, respectively. Through the model, we successfully screened 614 candidates with stable structures and suitable band gaps (1.34 eV ∼ 1.5 eV) from 1,569,248 perovskites. To verify the robustness and generalizability of our ensemble learning model and facilitate future researchers' exploration of high-performance perovskites. we selected the most widely used bromide perovskite and two new perovskites (sulfides and oxides) with excellent optoelectronic properties for further DFT validation. They all exhibit excellent structural stability, thermodynamic stability, and suitable band gaps, among which Cs2TlPBr6 and Pb2RhPaS6 also have direct band gaps, low surface work function, and low effective masses, which are expected to be candidates for high-performance perovskite solar cells. Our proposed ensemble learning model enables high-precision prediction of perovskite performance in an “inexpensive” way, greatly reducing the design cycle of perovskite solar cells and promoting the application and development of perovskite solar cells.