A Machine Learning-Assisted Approach to a Rapid and Reliable Screening for Mechanically Stable Perovskite-Based Materials

The present work is designed to discover new perovskite-based materials, which are expected to show high mechanical stability during their applications, using machine learning (ML) techniques, and based on the Pugh's criterion for distinguishing brittle and ductile behaviors. For this purpose, ML models to predict the moduli of materials, bulk (B) and shear (G), are built using their crystal structure and composition information. The ML process is initiated with the information of 5663 compounds, including composition, crystal structure and moduli, as listed in AFLOW database. Following a procedure of data characteristics, feature generation, feature processing, training, and testing, the ML models are constructed with acceptable accuracy (tenfold cross-validation R² score of 0.90 and 0.89 for B and G, respectively). The validation process of the models, which is conducted using the corresponding density functional theory calculations, reveals that these models are reliable to be employed in a large-scale screening process. Indeed, the B- and G-based ML models are incorporated in a screening process, and this is also conjugated with other screening criterions, to find out thermodynamically stable and formable perovskite-based materials with improved mechanical performance.