Perovskite solar cells (PSCs) are one of the most promising candidates for next-generation photovoltaics. Due to their exceptional increase in power conversion efficiencies (PCEs) over the last 15 years and their excellent suitability for tandem device architectures, the emerging technology has reached the brink of commercialization. This achievement was enabled by immense efforts in fast, manual prototyping based on a culture of apprenticeships, where each team develops fabrication processes customized for their specific equipment. While  this approach was very successful for improving the state-of-the-art PCEs, researchers should now consider focusing more on rigorousness of scientific reporting. The reason is that the field suffers from immense issues with reproducibility, leaving a huge potential for standardization and refining of routines. A simple indicator for the rigorousness of scientific reporting is the extent to which process parameters are controlled and provided in the description of experiments.

Leveraging extensive experience with modelling of perovskite processing[1], we perform a meta survey of the process parameters provided in published literature on perovskite solution processing. For conducting the analysis, ways of sampling representative meta data are developed. Previously, meta data was collected in an extensive effort establishing the perovskite open source database[2][3]. However, we found that the database prioritizes frequently reported over rarely reported parameters. Additionally, the acquisition of these data took between 5.000 to 10.000 of volunteer work hours. To shortcut the process of data acquisition, we leverage recent progress from the field of deep machine learning to collect data on perovskite solution processing on a representative sample of around 1.500 publications with minimal need for human intervention. However, this approach comes with novel challenges. Common deep learning algorithms do not have an understanding of the technology, such that the potential for misinterpretation and error is high. Therefore, the algorithm must be carefully assessed in comparison to human reading performance. The exact formulation of instructions, the choice of the specific model and its hyperparameters, as well as pre-training are critical indicators for increasing model performance. We find that deep leaning models can be on par with human reading performance (that is not error-free either), when given simple instructions. It is however an open question, if these models can be scaled to advanced meta data extraction projects like the perovskite database. Furthermore, a strategy on how to publish these models must be developed as the trained models could potentially be exploited to access copyright-protected content.

After conducting the analysis for each of the 10 process parameters, we identified as critically important for reproducing perovskite solution processing, we find that there is a great potential for improving the rigorousness of the description of experiments in perovskite solar cell fabrication. Furthermore, there are clear differences in the likelihood of certain parameters to be provided. For example, the spin speed (or the sheering velocity) of the perovskite solution deposition is provided in almost all cases, while the temperature of the atmosphere during processing is almost never reported. As consequence of our analysis, we propose that the specification of experiments in perovskite processing should be given more attention, in particular when reviewing and writing articles on perovskite solar cells (we do not exclude ourselves from this imperative). Because the technology is so successful that it is in the process of being passed on to industry, researchers now have the privilege of prioritizing rigorousness of reporting and process analysis over device optimization speed. This further opens up new opportunities for publishing studies on fundamental process physics and chemistry, differentiating technology development carried out at rapid pace in industry from research on fundamentals conduced by publicly funded research institutions.