Perovskite (PVK) materials have revolutionized solar cell research, achieving remarkable efficiency gains of over 26% in a short time span. A major factor behind this progress is the extensive research focused on absorber layer (AL) materials. The cell efficiency is largely influenced by the important physical properties of the AL, such as the bandgap (BG) and the energy levels of the valence band maximum (VBM) and conduction band minimum (CBM), which provide insights into the light absorption and the band offset with the charge transport layers. In this study, we implemented machine learning (ML) models to predict these three important parameters for AL. Using well-cleaned 551 experimental data from the perovskite database, we trained various ML algorithms, selecting CatBoost as the best performer for BG prediction with a test RMSE of an impressive 0.054 eV. The BG model showed very high accuracy when validated with 13 unseen experimentally reported data. This algorithm was further used to enhance the sets of ML models for CBM and VBM prediction, achieving RMSEs of 0.103 eV and 0.120 eV, respectively. For the refinement of the model’s performance, we have used Optuna for parameter tuning. Additionally, SHAP analysis provided valuable insights into the relative feature importance, aligning well with established theoretical principles for BG tuning through metal and halide ion mixing. These models represent the best-performing ML approach to date for predicting these parameters. This framework can dramatically enhance the search for optimal PVK compositions, reducing the need for resource-intensive experimentation.
Read full abstract