Revealing the Impact of CZTSe/CdS Interface Fluctuations on PV Device Performance through Big Data Analysis Assisted by Machine Learning Methods

Abstract

This work showcases the importance of developing suitable inspection and analysis methodologies with high statistical relevance data coupled with machine learning algorithms, for the detection, control, and understanding of small fluctuations in the scale-up of thin film photovoltaics to industrial sizes. To exhibit this methodology, this work investigates the effect of subtle inhomogeneities on the efficiency of thin film solar cells based on the Cu₂ZnSnSe₄/CdS interface using two large area samples subdivided in ≈400 individual solar cells. A large dataset obtained from Raman and photoluminescence spectroscopic techniques together with J–V optoelectronic data is generated to elucidate the impact of these inhomogeneities on the efficiency of the devices. Using a combination of statistical (spectral difference) and over 440 000 multivariate polynomial regressions through machine learning algorithms, it is revealed how the main limiting factor for device performance are subtle fluctuations in the nanostructure and surface defects of the CdS layer, rather than compositional fluctuations or defects in the kesterite absorber. It is estimated that the avoidance of these issues could result in an absolute increase in device efficiency of 2%. This could provide a potential avenue for further technology advancement within the kesterite community.