Machine Learning to Understand Solar Cell Performance

Abstract

This study contributes to the understanding of a perovskite solar cell (PSC) performance database which contains 7,026 data points gathered from research papers published between 2012 and 2020. The aim is to capture and analyse historical main patterns through machine learning technologies to generate models and heuristics for predicting cell performance. The dataset utilized has a total of 16 attributes, each of which contains several categories. The attribute for each device includes the structural parts (such as the electrodes, absorber layers, substrate, transport layers), cell architecture, type of module and band gap of the perovskite solar cell. By visualizing the data with Python, several factors for increasing the efficiency of PSC were identified. For instance, it was found that maximum device efficiency can be achieved by using polyimide or PSG as substrate, indium tin oxide (ITO) or `fluorine-doped tin oxide (FTO) as electrodes, and copper as electrode2. For the analysis, different machine learning models were constructed and tested, with random forest providing the best results in predicting perovskite solar cell efficiency. To identify the most important factors that influence solar power conversion efficiency (PCE), the sequential minimal optimization regression (SMOreg) model was also used. This is accomplished by looking at the attribute weights in the generated SMOreg model. The analysis given significantly enables the identification of variables and layers that may lead to improvements in PCE through device design, as well as emphasizing the involvement of various factors in the degradation of PSC.