Selecting an appropriate machine-learning model for perovskite solar cell datasets

Abstract

Abstract Utilizing artificial intelligent based algorithms in solving engineering problems is widely spread nowadays. Herein, this study provides a comprehensive and insightful analysis of the application of machine learning (ML) models to complex datasets in the field of solar cell power conversion efficiency (PCE). Mainly, perovskite solar cells generate three datasets, varying dataset size and complexity. Various popular regression models and hyperparameter tuning techniques are studied to guide researchers and practitioners looking to leverage machine learning methods for their data-driven projects. Specifically, four ML models were investigated; random forest (RF), gradient boosting (GBR), K-nearest neighbors (KNN), and linear regression (LR), while monitoring the ML model accuracy, complexity, computational cost, and time as evaluating parameters. Inputs' importance and contribution were examined for the three datasets, recording a dominating effect for the electron transport layer's (ETL) doping as the main controlling parameter in tuning the cell's overall PCE. For the first dataset, ETL doping recorded 93.6%, as the main contributor to the cell PCE, reducing to 79.0% in the third dataset.

Keywords

• Random forest
• Gradient boosting
• K-nearest neighbors (KNN)
• Linear regression (LR)
• Hyperparameter tuning
• Power conversion efficiency