Process Understanding of Scalable Perovskite Thin-Film Formation Through Unsupervised Machine Learning of In Situ Luminescence Data

High performance perovskite solar cells (PSCs) at laboratory scale show that hybrid metal-halide perovskite semiconductors are a promising absorber material class for the next generation thin-film solar cells. However, transferring the processes to scalable methods that enable large-scale processing remains a key challenge hindering the commercialization of the technology. To enable large-scale and high-throughput production of PSCs, the formation and morphology of perovskite thin-films fabricated by scalable deposition methods needs to be optimized. The perovskite formation process includes the entangled process phases of drying, nucleation and crystal growth, which are crucial for the quality of the perovskite thin-film and must be controlled to obtain thin-films of high optoelectronic quality. Consequently, scaling the technology is complex and demands enhanced understanding of the highly intricate formation process. To improve this understanding and the reproducibility of the scalable thin-film formation process, data-driven machine learning (ML) methods can be employed to accelerate research and process control of large-area perovskite thin-films. For this, a unique, labelled <i>in situ</i> photoluminescence (PL) dataset was generated. The dataset contains <i>in situ</i> PL data of more than 1,100 PSCs captured during the vacuum-assisted annealing of the blade-coated perovskite thin-films with an in-house-built imaging setup. To generate the ML-based in-line process monitoring dataset, all solar cells were fabricated at the exact same conditions, e.g. layer stack and precursor materials.<br/>In this work, we present unsupervised ML for process understanding of perovskite thin-film formation by employing k-means clustering to our unique <i>in situ</i> PL dataset. We show the benefit of acquiring <i>in situ</i> PL data during the perovskite formation over <i>ex situ</i> PL data by demonstrating the correlation between expert-chosen <i>in situ</i> PL features and the power conversion efficiency (PCE) of the corresponding PSCs. Next, we show that k-means clustering creates <i>in situ</i> PL clusters that correlate with the performance of the final PSC without prior data encoding assumptions made by a human expert. The correlations display that the <i>in situ</i> PL data contains information about the quality of the perovskite thin-film and, consequently, about the performance of the solar cell. Furthermore, we identify detrimental process mechanisms during the formation of the perovskite thin-film using the data science approach. Next to the clusters’ performance correlation, differences in the spatial distributions of PL data assigned to the different clusters are identified and reveal substrate areas with adverse thin-film properties. Finally, when applying the trained model to previously unseen data, we find that k-means clustering allows for reasonably good prediction of solar cell performance even though clustering is typically used mainly to explore datasets. In summary, we demonstrate that ML-based in-line processing monitoring of the formation of perovskite thin-films holds great potential and can accelerate the successful commercialization of perovskite thin-film PV.