Perovskite-based photovoltaic devices show promising efficiency, nearing that of crystalline silicon. However, scaling up device area remains a challenge, particularly in achieving homogeneous perovskite crystallization over large substrates using solution-coating methods like slot-die coating. Slot-die coating, while advantageous for high-throughput and efficient material use, can lead to inhomogeneities (thickness variations, cracks, pinholes) due to gas-knife-induced mechanical deformation and drying. Optimizing coating parameters is crucial for homogeneity, but rapid quantification of large-area film homogeneity is currently lacking. Existing methods like optical spectroscopy, XRD, and stylus profilometry are limited in speed or destructiveness for large samples. Manual visual inspection, while fast, is subjective. This research addresses this gap by applying machine vision to automate and quantify the visual inspection of perovskite films, building upon previous work in machine vision for silicon photovoltaics and thin-film analysis. The goal is to develop a machine vision tool that extracts quantitative information relevant to optimizing large-area perovskite film deposition, specifically using gas-knife assisted slot-die coating. This involves building comprehensive maps linking coating parameters (wet film thickness, gas-knife speed) to perovskite film properties and using an optical model to predict device current density. This allows identifying optimal process conditions that maximize throughput, film quality, and device performance.

The authors reviewed existing literature on machine vision techniques for inspecting silicon photovoltaics, quantifying thin-film thickness, and detecting morphological defects. They cite studies showing the effectiveness of machine vision in detecting defects in silicon photovoltaic devices and quantifying the thickness of various thin films. The literature also highlights the use of machine learning in optimizing perovskite compositions, but less emphasis on using machine vision for optimizing perovskite processing strategies. One notable exception is the use of a convolutional neural network to detect bulk perovskite crystal formation during antisolvent crystallization optimization. This existing work provides a foundation for the development of specialized machine vision tools tailored to analyze perovskite film morphologies and ultimately optimize large-area perovskite film deposition techniques.

The study used gas-knife-assisted slot-die coating to deposit CS0.16FA0.84Pb(I0.88Br0.12)3 films on SnO2-coated ITO glass substrates. The process parameters varied were gas-knife speed and wet precursor film thickness. Perovskite films were characterized using white-light photography. A convolutional neural network (CNN), based on the VGG16 architecture, was trained to segment images into areas of full, partial, and no coverage. The CNN was further extended to estimate perovskite film thickness with pixel resolution (10 µm x 10 µm) using a calibration curve derived from profilometry measurements of spin-coated reference samples. An unsupervised model, based on the work of Jeon et al., was used to detect and quantify morphological defects (cracks, pinholes) in the fully covered areas. The extracted perovskite film properties (thickness, defects) were linked to photovoltaic device performance using a physics-based optical model employing the transfer matrix formalism. This optical model predicts an upper bound on device photocurrent density, assuming 100% internal quantum efficiency (IQE). The model uses optical indices and thicknesses of each layer in the device stack (obtained via ellipsometry measurements). To validate the optical model, photovoltaic cells with spin-coated perovskite films of varying thicknesses (160, 250, 350, 550 nm) were fabricated and characterized. Finally, the integrated workflow (PerovskiteVision and the optical model) was used to predict the current density of perovskite photovoltaic modules.

PerovskiteVision successfully quantified substrate coverage, film thickness, and defect density from optical images. The CNN model accurately segmented images into regions of full, partial, and no coverage. The thickness estimation model, calibrated using profilometry data, provided pixel-by-pixel thickness maps. The defect detection model identified morphological defects, enabling the calculation of defect metrics. The study identified optimal slot-die coating parameters (gas-knife speed and wet film thickness) that maximized substrate coverage, minimized defect density, and yielded high predicted current density. The optical model accurately predicted device current density for devices with spin-coated absorbers, except for the thinnest absorbers (160 nm), likely due to PbI2 impurities. When applied to devices with slot-die coated absorbers, the model provided accurate predictions for some devices, but overestimated the current density for others, highlighting limitations in not accounting for all non-idealities. PerovskiteVision effectively predicted the current density of a hypothetical perovskite photovoltaic module, demonstrating its potential for module design and optimization. The spatially resolved predictions highlight how defects in individual cells can limit overall module performance.

This work successfully demonstrates the utility of machine vision in accelerating the optimization of large-area perovskite thin-film deposition. PerovskiteVision provides a fast, reliable, and non-destructive method for characterizing film homogeneity, which is crucial for large-scale manufacturing. The integration of the optical model enables direct linking of process parameters to device performance predictions. The ability to identify optimal process conditions that maximize both performance and throughput is significant for the advancement of perovskite photovoltaics. While the optical model showed limitations in predicting the current density of devices with very thin absorbers or those with significant non-idealities, the overall approach significantly improves the efficiency of process optimization compared to traditional methods. Future work could focus on improving the accuracy of the optical model by incorporating additional measurements (e.g., ellipsometry) to account for sample-to-sample variations in optical properties. Photoluminescence imaging could also provide further insights into opto-electronic quality. Adapting the approach to larger substrates would enhance its applicability to industrial quality control.

PerovskiteVision offers a significant advancement in the characterization and optimization of large-area perovskite films. By integrating machine vision with an optical model, this tool enables rapid, non-destructive quantification of key film properties and prediction of device performance. This accelerates the multi-objective optimization of deposition processes, bridging the gap between research and large-scale manufacturing. Future directions include enhancing the optical model, applying the method to larger substrates, and integrating it into online feedback systems for manufacturing or autonomous laboratories.

The optical model used in this study assumes spatially homogeneous optical properties within the perovskite film, which may not always be true. The presence of PbI2 impurities in the perovskite films, particularly in thinner films, affects the accuracy of the current density predictions. The model does not account for all types of non-idealities affecting device performance (e.g., open-circuit voltage, fill factor), leading to overestimation of current density in some cases. The current version of PerovskiteVision is limited to 5cm x 5cm substrates; adapting it to larger substrates would enhance its industrial applicability.