Advancing scalability & process benchmarking in mechanical exfoliation of superconducting 2D materials

Abstract

Mechanical exfoliation remains the most accessible route to atomically pristine, large-area, single-crystal 2D flakes for superconducting quantum devices, multidimensional photodetectors, and biosensing systems. However, the process is slow, operator-dependent, prone to variability, and often guided by heuristics lacking practical empirical validation. This work introduces an end-to-end framework to enable the shift for mechanical exfoliation procedures from a manual, artisanal procedure into a measurable, optimizable process by presenting the following: (i) a low-cost, open, and electronically controlled instrument that enforces adhesive peel kinematics and normalizes tape-removal speed across angles; (ii) a GPU-accelerated pipeline using a machine learning segmentation model to perform optical-classification for high-throughput 2D material flake thickness mapping; and (iii) a systematic study isolating peel angle and speed to quantify yield-residue trade-offs using computer vision techniques to normalize comparisons. The instrument, built from off-the-shelf parts, delivers precise angle control ( 0◦ to 120◦) and verified speed accuracy with ±0.7% relative error when incrementally tested within a range of 1 to 5000 μm/s. The sample image classification software extends existing density-based clustering methods with automatic background masking and vectorized GPU kernels. Benchmark comparisons show that this implementation processes over 200× more pixel data with a 60× reduction in processing time relative to the original software while preserving ∼95% pixel-wise accuracy, enabling full-resolution analysis of thousands of microscope images on consumer hardware. Using these tools, 72 controlled depositions (6 angles × 4 speeds, ? = 3 each) were executed and 209,952 sample images were classified, normalizing outcomes by tape-coverage to decouple upstream variability. Results suggested more nuanced interactions of speed and angle parameters and challenged existing conventions that slow speeds are optimal for material yields. While slowest peels maximized single-layer coverage overall, they also drove the highest amounts of contamination residue. Several higher-speed/angle combinations achieved ≥80% of the peak single-layer yield with ≤30% of the residue, revealing tunable operating points that favor cleanliness without severely sacrificing monolayer area. Taken together, the open instrument, software, and data-driven optimization testing framework provide practical guidance for reproducible exfoliation, and a scalable baseline for benchmarking materials, substrates, and future automation.