Using scalable computer vision to automate high-throughput semiconductor characterization

The rapid acceleration of high-throughput (HT) synthesis across diverse material systems has created a throughput bottleneck in characterization, particularly for optoelectronic semiconductors where properties such as band gap and environmental stability are vital. Inkjetting and drop-casting enable 10^4 samples per hour, but conventional characterization is slow, manual, and optimized for uniform thin films rather than droplet-like deposits with variable morphology. Existing semi-automated or HT methods often assume fixed sample geometry and serial, hard-coded measurement positions, limiting scalability and generalizability. To bridge this gap, the authors develop adaptive, computer vision-based, parallel autocharacterization tools that segment arbitrarily many non-uniform samples and compute three key properties—composition, optical band gap, and environmental degradation—within minutes. The contributions include: (1) scalable sample detection and segmentation on hyperspectral/RGB data, (2) composition mapping tied to synthesis pump speeds and tool paths, (3) automated direct band gap extraction from hyperspectral reflectance, and (4) automated stability assessment from RGB time-series, demonstrated on N=201 FA1−xMAxPbI3 samples (0 ≤ x ≤ 1).

Prior work has automated aspects of band gap computation from pre-collected optical data for single samples (e.g., Escobedo et al.) and expanded to HT optical characterization of up to 24 samples per batch with static, serial measurements (Surmiak et al., Reinhardt et al.). Other platforms achieved large-scale synthesis/characterization via microplates or robotic systems (Langner et al., Wang et al., Du et al.) but relied on rigid geometries. Wu et al. demonstrated end-to-end automation for optical properties of molecules in solution; however, methods do not generalize to deposited materials with variable morphologies. In materials imaging, computer vision has enabled segmentation, denoising, and scalable analysis for microstructures (Park et al., Chowdhury et al., Li et al., Wang et al., Neshatavar et al., Tung et al., Jain et al.). These advances motivate integrating computer vision into semiconductor characterization to parallelize measurements across numerous non-uniform samples without sacrificing speed or accuracy.

Computer vision segmentation and mapping: Hyperspectral datacubes (e.g., 900×800×300 wavelengths) are processed with a pipeline of cropping, grayscale conversion, binarization, morphological gradient, erosion, median blur, distance transform, feature labeling, and watershed segmentation to identify each deposit island. Features are smoothed, pruned by size thresholds, and a boolean mask maps each sample’s pixels (X,Y) to its reflectance spectra R(λ), enabling parallel association Φ=(X,Y,R(λ)). The approach scales beyond 80 samples in parallel and is agnostic to sample size/geometry (Algorithm 1). Composition mapping: HT printing deposits FA and MA precursors with time-varying pump speeds ωFA and ωMA while the print head rasters in a serpentine path. G-code tool path coordinates with timestamps are overlaid on segmented samples to assign each deposit a deposition time window Δt=[ta,tb]. Pump speed time series from the microcontroller are timestamp-matched to deposits, and composition x (MA fraction) is computed by integrating x(t)=∫ta^tb ωMA(t)/(ωMA(t)+ωFA(t)) dt. Validation is via XRD peak shifts (MA-rich shift to higher 2θ) and XPS C=N vs C−N peak evolution across the gradient. Band gap automation: For each deposit, the spatial median reflectance R(λ) (380–1020 nm) is transformed using the Kubelka–Munk function to a Tauc curve F(R(λ))·(hv)^γ = B(hv−Eg), with γ=1/2 for direct gaps. Tauc curves are recursively segmented until near-linear segments achieve R^2 ≥ 0.990. Adjacent segment pairs define candidate linear fits between Tauc peaks. A bounded RMSE minimization (between the regression x-intercept lower bound and the Tauc peak location minus half-peak width) selects the best fit; Eg is the x-intercept. Degradation detection: Samples are aged for 2 h at ~0.5 suns, 34.5±0.5 °C, 40%±1% RH. RGB images are captured every 30 s. Color calibration uses a reference color chart and 3D thin-plate spline warping to CIE 1931 (2° observer, D50 illuminant) to correct illumination-induced distortions. For each deposit, the average color R(t;X,Y) is computed over time; the instability metric Ic integrates the color change Ic=∫0^T [R(t;X,Y)−R(0;X,Y)] dt across r,g,b channels. Ground truth degradation is defined by expert-assessed band gap deviation >0.02 eV pre- vs post-aging. Performance is evaluated via precision-recall AUC and accuracy. Experimental details: Substrates are cleaned (DI/Hellmanex, DI, IPA) and processed in N2 glovebox (<10 ppm). FAPbI3 and MAPbI3 0.6 M precursors are prepared in DMF:DMSO (4:1). Inkjet HT printer deposits ~70–80 compositions per batch in 16.5 s, with pinch-valve dropletization and mixing, followed by 150 °C anneal for 15 min. Hyperspectral imaging (Resonon Pika L) and RGB imaging supply data for autocharacterization.

• Parallel computer vision segmentation mapped pixels to spectra for N=201 FA1−xMAxPbI3 samples and scaled beyond 80 samples in parallel. - Composition mapping validated: XRD plane (012) near 2θ≈31.5° shifted by Δ2θ=0.16° from FA-rich to MA-rich; XPS showed decreasing C=N peak (~400 eV) intensity with increasing MA content, confirming the structural and elemental gradient. - Band gap autocharacterization vs domain expert (N=201 across 3 batches): R^2=0.975 linear correlation; 98.5% accuracy within ±0.02 eV; systematic slight underprediction noted. Throughput: 6 min for 200 samples vs ~510 min manually (≈85x faster). Derived band gaps consistent with literature: FAPbI3 ≈1.46 eV and MAPbI3 ≈1.55 eV; generated ultra-high-resolution trend with 120 unique compositions (over 13x increase in compositional resolution vs prior reports). - Degradation autocharacterization: precision-recall AUC=0.853; maximal accuracy=96.9% relative to ground truth classification (defined by >0.02 eV Eg shift). Throughput: full computation 20 min for 200 samples using 48,000 images over 2 h. Identified high degradation for FA-rich compositions (x≈0.0–0.15) consistent with α-FAPbI3 → δ/β-FAPbI3 pathway; optimal low-degradation region near x≈0.40. Achieved >10x increase in compositional and ~40x increase in temporal resolution, totaling ~436x more unique data points vs prior work. - Overall, autocharacterization enables 10^3 samples·h−1-scale measurements, narrowing the gap to HT synthesis (10^4 samples·h−1).

The presented autocharacterization suite addresses the characterization bottleneck in HT materials discovery by enabling parallel, adaptive measurements on non-uniform deposited samples. Computer vision segmentation decouples measurement from rigid geometries, allowing hyperspectral and RGB analyses across hundreds of samples simultaneously. The band gap tool delivers expert-level accuracy with major throughput gains, yielding fine-grained composition–Eg trends that agree with literature endpoints and enable discovery at unprecedented resolution. The degradation tool leverages colorimetric proxies under controlled aging, achieving high classification performance and rapid turnaround, and reproduces known degradation regimes in FA–MA perovskites. Together, these advances demonstrate that integrating scalable computer vision with HT synthesis can synchronize or approach synthesis rates, thereby accelerating exploration of complex composition–property landscapes in semiconductors.

This work introduces scalable computer vision-driven autocharacterization tools that automatically segment and map composition, compute direct band gaps from hyperspectral reflectance, and quantify environmental degradation from RGB time series for FA1−xMAxPbI3 perovskites. The methods achieve expert-level accuracy (band gap 98.5% within ±0.02 eV; degradation accuracy 96.9%) with large speedups (band gap: 6 min/200 samples; degradation: 20 min/200 samples), enabling ultra-high resolution composition and stability trends and narrowing the throughput gap with HT synthesis. Future work will extend band gap analysis to multi-phase materials featuring multiple band edges, further generalize the approach to broader material classes, and enhance robustness via tighter control of environmental and imaging conditions.

• Current band gap autocharacterization is limited to single-phase materials and direct band gaps; multiphase samples with multiple band edges are not yet handled. - Degradation measurements can be affected by experimental variabilities (vibrations, lighting fluctuations) despite color calibration; stringent control of chamber conditions is necessary for optimal reproducibility. - Systematic underprediction in automated band gap fitting relative to expert indicates potential model bias or fitting window limitations that could be refined.