A machine vision tool for facilitating the optimization of large-area perovskite photovoltaics

Perovskite-based photovoltaic devices are nearing crystalline silicon performance but at much smaller device areas, creating a need to scale fabrication while maintaining film homogeneity. Large-area deposition via slot-die coating with gas-knife quenching enables high-throughput processing but introduces challenges such as thickness variations, cracks, and pinholes due to mechanical deformation and drying inhomogeneities. Rapid, non-destructive, and quantitative methods to assess large-area perovskite film homogeneity are lacking; existing techniques (optical spectra, XRD, profilometry) are limited by small sampling area, destructiveness, or slow acquisition. Manual visual inspection, while fast, is subjective and unsuitable for rigorous process optimization. Motivated by this gap, the study proposes automating and quantifying visual inspection through machine vision to extract spatially resolved metrics pertinent to optimizing large-area perovskite deposition processes.

Prior work has applied machine vision to silicon photovoltaics for detecting scratches, cracks, dirt, and defects from color or electroluminescence images and surveyed inspection strategies. Machine vision has also quantified thin-film thickness and morphological defects (pinholes, cracks, dewetting) and enabled rapid, non-contact, spatially resolved quality control over large areas. Machine learning has been used to optimize perovskite compositions for performance and stability and to monitor crystallization (e.g., CNN detecting bulk perovskite crystal formation for antisolvent crystallization optimization). However, there has been limited use of machine vision specifically to optimize perovskite processing and large-area deposition. This work tailors machine vision tools to analyze perovskite film morphology, linking process parameters to film properties and device-relevant predictions.

Process and samples: Gas-knife-assisted slot-die coating was used to deposit Cs0.16FA0.84Pb(I0.88Br0.12)3 perovskite films on 5 cm × 5 cm SnO2-coated ITO glass. The slot-die process dispenses a wet precursor film followed by nitrogen gas quenching to trigger crystallization. Coating parameters varied included gas-knife speed (5–33 mm·s−1) and theoretical wet film thickness (controlled via dispense rate, coating speed, and substrate width; typical coating speed 28 mm·s−1, dispense 100 µL·min−1, coating gap 100 µm, 1–4 µm wet thickness). Coating was conducted in an enclosed environment (20–30% RH), with substrate temperature 25–100 °C. Imaging: Bright-field transmission photographs were taken on a luminous background (Olympus DP70, 4080×3072 px, 1 s exposure, ISO 200, 40× magnification). 504 slot-die coated samples were imaged. Profilometry (Bruker Dektak) provided thickness references; for calibration, 9 positions per 5×5 cm substrate were measured. Substrate coverage segmentation: A CNN based on VGG16 architecture was trained to classify 50×50 px image patches into covered, partially covered, or uncovered. Training used 42 manually annotated images (sliding window size 50 px with 45 px overlap). Background removal identified coated area; coverage ratio was defined as covered area divided by total coated area. Data split: 70% train, 20% validation, 10% test with five-fold cross-validation. Hyperparameters: batch size 100, initial learning rate 0.001, Adam optimizer. Performance: 95.28% validation accuracy (peak at 9th iteration), 88.69% test accuracy. Thickness estimation: Film thickness within fully covered regions was estimated pixel-by-pixel (10 µm × 10 µm per pixel) by mapping image color (gray/RGB) to profilometry-derived thickness via non-linear regression. Calibration used 135 data points from 11–15 spin-coated reference films spanning ~160–550 nm (typical slot-die range). Extrapolation up to 800 nm was applied where necessary but not used for current-density predictions. Thickness maps were generated for all 504 slot-die samples. Defect detection: An unsupervised pixel-level defect detection method derived from Jeon et al. combined Canny edge detection with an adversarial image-to-frequency transform to identify morphological defects (cracks, pinholes) within fully covered areas, producing binary defect maps and enabling metrics such as defect area fraction. Optical modeling of Jsc: A physics-based transfer matrix optical model (Stanford TransferMatrix_VaryThickness) predicted an upper bound of photocurrent density (Jsc) assuming IQE = 100%. Layer stack: glass/ITO/SnO2/perovskite/PTAA/Au. Optical constants for each layer were extracted by variable-angle spectroscopic ellipsometry (50°, 60°, 70°, 0.6–4 eV) on representative stacks. Perovskite optical properties were taken from 350 nm spin-coated films and used for all films. The model was applied per pixel to thickness maps; uncovered/partially covered pixels contributed zero current. Validation devices with spin-coated perovskite thicknesses (160, 250, 350, 550 nm) were fabricated to compare measured Jsc to model predictions. Device fabrication and testing: ITO glass (~7 Ω/sq) was cleaned (acetone, IPA, DI water) and baked (100 °C overnight). SnO2 compact layer was spin-coated from a nanoparticle solution (3% w) and annealed (80 °C, 1 min). Spin-coated perovskites for calibration used anti-solvent quenching (200/1000/6000 rpm with 700 µL chlorobenzene drop) and annealed (100 °C, 1 h, N2). Slot-die parameters as above with gas flow 110 L·min−1, gas-knife 3 mm above substrate. JV measurements used AM1.5G (Oriel 92190, 1600 W Xe lamp), Keithley 2602A; reverse scan 1.2 to −0.2 V; device area 0.33 cm²; five repeated measurements per device. Module analysis: For a hypothetical 4×4 cm module comprising eight series-connected sub-cells, PerovskiteVision-generated Jsc maps were analyzed over the designated module area to compute average and limiting cell currents and infer module-limiting performance due to defects.

• PerovskiteVision quantified substrate coverage, pixel-resolved thickness (10 µm × 10 µm), and morphological defects from white-light photographs of 5×5 cm perovskite films (504 samples analyzed).
• Coverage segmentation achieved 95.28% validation accuracy and 88.69% test accuracy using a VGG16-based CNN.
• Thickness estimation calibrated on 11 spin-coated references (160–550 nm) enabled generation of thickness maps across the dataset; extrapolated values (up to 800 nm) were not used for current predictions.
• Defect detection via unsupervised image-to-frequency transform identified cracks and pinholes; defect metrics such as area fraction were computed. Maximum observed defect area was 1.5% for thick wet film at low gas-knife speed (16 mm·s−1, 1.8 µm wet thickness).
• Process–outcome mapping: Lower gas-knife speeds (15–19 mm·s−1) yielded >60% coverage for both thin and thick wet films; intermediate speeds (21–23 mm·s−1) reduced coverage and increased mean thickness; high speed (33 mm·s−1) increased mean thickness and defects.
• Three favorable samples with >60% coverage and low defect density were identified: A (15 mm·s−1; 2.0 µm wet), B (28 mm·s−1; 1.2 µm wet), C (15 mm·s−1; 0.45 µm wet).
• Optical model validation on spin-coated devices showed good agreement except for 160 nm absorbers, where measured Jsc fell below predictions, attributed to PbI2 impurities and sensitivity of thin films to extinction coefficient variations.
• Using the optical model per pixel, predicted Jsc maps linked process parameters to device performance. Sample B (28 mm·s−1, 1.2 µm wet) provided the highest simulated Jsc ≈ 22.7 mA·cm−2 with comparable defect density to A and C, and higher process speed.
• For slot-die coated devices fabricated under sample B conditions (~550 nm absorber), spatial Jsc predictions correlated well with measurements for devices 3–5; devices 2 and 6 were overestimated due to unmodeled non-idealities.
• Module-level prediction demonstrated that a single cell with a large pinhole (cell #4) limited overall module current due to series connection, illustrating pre-fabrication layout optimization using PerovskiteVision.

The study addresses the critical need for rapid, non-destructive, and quantitative assessment of large-area perovskite film homogeneity, a bottleneck for scaling slot-die coated modules. By automating visual inspection, PerovskiteVision provides pixel-level maps of coverage, thickness, and defects from a single photograph, enabling immediate, objective feedback compared to subjective manual inspection or slow, destructive profilometry. Integrating a physics-based optical model translates morphology (thickness, coverage) into upper-bound photocurrent density predictions, facilitating multi-objective optimization of process speed, yield (defect minimization), and device performance. The approach successfully identified favorable gas-knife speed and wet thickness conditions (e.g., 28 mm·s−1, 1.2 µm wet) and accurately predicted Jsc for several devices, while revealing limitations where unmodeled factors (e.g., impurities, recombination, series resistance impacting Voc/FF) affect performance. Module-level analyses demonstrate how spatial inhomogeneities dictate series-connected device performance, enabling informed selection of device regions or redesign prior to fabrication. Overall, the findings show that machine vision can accelerate essential characterization steps and streamline the optimization of large-area perovskite deposition processes.

PerovskiteVision converts white-light images of large-area perovskite films into quantitative, pixel-resolved maps of substrate coverage, thickness, and defect density, and—when combined with a transfer-matrix optical model—into spatially resolved Jsc predictions. This workflow enables rapid, non-contact, and holistic film assessment, supports creation of process parameter–outcome maps, and facilitates multi-objective optimization of gas-knife-assisted slot-die coating. The tool effectively guided selection of process parameters yielding high predicted photocurrent and low defect density and provided accurate Jsc predictions for several slot-die devices. Future work could improve accuracy by incorporating sample-specific optical constants (e.g., per-film ellipsometry), integrating photoluminescence or hyperspectral imaging to assess opto-electronic quality, and adapting to reflection imaging for opaque substrates (relevant to perovskite/silicon tandems). Integration into on-line manufacturing feedback and autonomous laboratories, coupled with multi-objective optimization algorithms, can further accelerate process development and quality control.

• Optical model assumes homogeneous optical properties equal to those of 350 nm spin-coated films and IQE = 100%, potentially overestimating Jsc, particularly for thin absorbers (e.g., 160 nm) and in the presence of impurities (PbI2) or electronic non-idealities.
• Device performance factors affecting Voc and fill factor are not modeled; thus, predictions reflect an upper bound on Jsc only.
• Thickness calibration was based on a limited set of spin-coated references (160–550 nm) with extrapolation up to 800 nm; extrapolated regions were excluded from Jsc predictions but may affect thickness map accuracy at extremes.
• Imaging modality relies on bright-field transmission; applicability to opaque substrates requires adaptation (reflection, PL, or hyperspectral imaging).
• CNN coverage segmentation, while accurate on validation (95.28%), showed lower test accuracy (88.69%); performance may vary with imaging conditions and sample types and requires initial labeled data.
• Defect detection is unsupervised and may miss subtle or electrically relevant defects not captured by morphological features alone.