Discovering novel lead-free mixed cation hybrid halide perovskites via machine learning†

In our recent study (S. Alidoust, F. Jamalinabijan and A. Tekin, ACS Appl. Energy Mater., 2024, 7, 785–798), a thorough computational screening using density functional theory (DFT) was conducted on mixed cation halide perovskites with a general formula of AA′BX₃, aiming to identify promising lead-free candidates. Employment of 23 A/A′-cations, 29 B-ions, and 4 X-anions yielded approximately 29 000 possible perovskite combinations. However, while modern high-throughput DFT frameworks can handle large-scale calculations, treating the entire configurational space of 29 000 possible perovskite combinations remains computationally demanding. Leveraging machine learning (ML) approaches could provide a more efficient alternative for capturing this complexity. Therefore, by using two empirical criteria known as octahedral and tolerance factors, this huge number was narrowed to nearly 2700, and the corresponding decomposition energy and band gap calculations were performed for each one of them. However, the remaining nearly 26 300 perovskites, though not selected by the empirical criteria, could still hold valuable and potentially promising candidates. Therefore, an ML model has been trained on the DFT-calculated subset, which has been increased to 4181 to achieve molecular and elemental homogeneity in these data sets to predict and identify promising perovskites within the unexamined portion of the dataset. Remarkably, the ML approach identified 930 promising perovskites satisfying both the decomposition energy (≤0.025 eV per atom) and band gap (1.0 ≤ gap ≤ 2.0 eV) criteria. Among these, 20 perovskites were selected for further validation through DFT calculations, and a very nice agreement has been obtained between the predicted and calculated decomposition energy and band gap values. These findings highlight the effectiveness of ML in accelerating the discovery of materials with specific desirable properties.