Perovskite solar cells are currently in the headlines, as they show high efficiencies in the conversion of solar energy into electrical energy in the laboratory. In combination with the proven silicon solar cells, more than a few people see this as the future generation of photovoltaics. Researchers at the Karlsruhe Institute of Technology (KIT) have now shown that machine learning (ML) can be a crucial tool to improve the measurement technology required for the commercial production of perovskite solar cells.

Photovoltaics is one of the key technologies for a low-carbon energy supply. Solar cells made of perovskite semiconductor layers already exhibit very high efficiencies and can also be produced cost-effectively. Another advantage: this technology enables particularly thin and flexibly designed modules. “Perovskite photovoltaics is on the threshold of commercialization. However, challenges remain in terms of long-term stability and upscaling to large areas,” says Professor Ulrich Wilhelm Paetzold, who conducts research at the KIT Institute for Microstructure Technology and the KIT Light Technology Institute (LTI). ‘In our study, we show that machine learning is crucial to improve the monitoring of perovskite thin-film formation required for industrial production,’ the physicist says. Using deep learning – a machine learning method that utilizes neural networks – the KIT researchers were able to quickly and accurately predict the material properties and efficiency of solar cells, even beyond the laboratory scale.

“Based on measurement data collected during production, machine learning can be used to identify process errors before the solar cells are completed. Additional inspection methods are not necessary,” says Felix Laufer, research associate at LTI and first author of the study. ”The speed and power of this method significantly improves data analysis. It can be used to solve tasks that would otherwise be difficult to accomplish.” The study of a novel data set documenting the formation of perovskite thin films enables deep learning to precisely map process data to target variables such as energy conversion efficiency.

“Perovskite photovoltaics has the potential to revolutionize the photovoltaic market,” says Paetzold, who heads the Next Generation Photovoltaics department at LTI. ”We are showing how process fluctuations can be quantitatively analyzed by extending the characterization methods with machine learning techniques. This ensures high material quality and layer homogeneity over large areas and many batches. This is a decisive step towards industrial applicability,” emphasizes the scientist.

Video