Performance prediction and optimization of perovskite solar cells based on the Bayesian approach

The development of perovskite solar cells (PSCs) has received much attention in recent years, but material selection schemes based on trial-and-error methods have made the enhancement of perovskite solar cell performance a huge challenge. Here, we propose a machine learning framework to predict the Photoelectric conversion efficiency (PCE) of PSCs with high speed and accuracy and use a Bayesian algorithm to inverse predict the optimal values of the underlying parameters (band gap, thickness of each layer, defect density, etc.) of PSC devices. Four different machine learning models, including the Extremely Randomized Trees (ExtRa Trees), k-nearest neighbor (KNN), light gradient boosting machine (LGBM), and random forest regression (RF) were adopted for training, with the ExtRa Trees model giving the best prediction performance (R2 = 0.71). In addition interpretable machine learning was applied to extract the performance-structure relationship for the PSCs. A Bayesian approach was used to rapidly derive optimal values for the cell device parameters at a PCE of 23%, which greatly accelerated the PSC manufacturing process. This work demonstrates the great potential of machine learning to accelerate the development of PSCs.