Building a device from a two-dimensional crystal starts with a person, or increasingly a neural network, scanning optical microscope images for a flake of exactly the right thickness. Thickness matters because many properties change abruptly at a critical layer count. Black phosphorus has a bandgap of 0.3 eV in bulk and 2.0 eV as a monolayer, and ferroelectric order in conventional films collapses below a threshold thickness. Real flakes are rarely clean. They sit under patches of tape glue and polymer residue, overlap other flakes, and carry thicker terraces of the same crystal. Current automatic classifiers read thickness from optical contrast, meaning how much the flake changes the reflected color relative to the bare substrate. They break down exactly where these obstructions appear. MaskTerial names crowded fields and small flakes as its failure modes. CLIFF only ever sees clean, pre-cropped flakes.
We propose to treat residue and overlaps the way occluded person re-identification treats pedestrians hidden behind cars, with one deliberate reversal. Re-identification models "repair" hidden body parts so they can still produce an identity. A fabrication tool should do the opposite. It should output a per-pixel thickness map together with a per-pixel trust map, and then certify only the contiguous regions whose thickness it can actually see. The trust map is learned from physically rendered occlusions, so it comes with exact labels at no annotation cost. The tool's success is measured by how rarely it certifies a wrong-thickness region that someone then spends a lithography run on.
Core Claim
A thin residue film laid over a monolayer can, for some film thicknesses, produce nearly the same reflected color as a clean bilayer. We call this contrast aliasing. Where it happens, no classifier that looks at one pixel's color can tell the two cases apart, and density-based uncertainty will not flag the pixel either, because the color looks like a perfectly normal bilayer. Spatial context can still separate them. Residue films vary smoothly and form blobs with irregular outlines. True terrace steps are sharp and tend to follow crystal edges.
Our claim has two parts. The first is that a model trained on physically rendered occlusions, with an explicit visibility head supervised by the renderer, will cut the false-certification rate for target-thickness regions well below that of per-pixel contrast classifiers and of MaskTerial's uncertainty. The second is that it will do so without losing accuracy on clean flakes. The whole proposal depends on aliasing being real for common substrate, residue and material combinations. That is a question the transfer-matrix physics can answer before any training, so it is our first decisive check.
Relation to Prior Work
The method borrows two ideas from occluded re-identification. Learning to Know Where to See uses a pose estimator only during training, to supervise a head that predicts continuous visibility scores from features. At test time it runs at 404 frames per second with 62.2% Rank-1 on Occluded-DukeMTMC. We reuse that training-only supervision pattern, but our supervision is exact and comes from physics rather than from a noisy external estimator. Pose-Guided Visible Part Matching restricts matching to visible regions, which is the principle behind our certification rule.
We deliberately do not adopt the repair-style mechanisms of HOReID or the attention pruning in MHSA-Net. HOReID uses graph neighbors to fill in occluded part features. MHSA-Net drives occluder attention to zero inside a single global descriptor. Both are useful when a global identity must be produced. For fabrication, a hallucinated thickness under residue is precisely the error that wastes a device.
On the materials side, MaskTerial supplies the rendering pipeline we extend. It generated 300,000 synthetic images from 35,000 real flake outlines using transfer-matrix modeling, which computes reflected color for a stack of thin films. It then segments flakes with Mask2Former and classifies thickness with a mixture model plus deterministic uncertainty. Its synthetic images contain no occluders, and its thickness assignment is per instance rather than per region with abstention.
Physically informed machine-learning algorithms constrains our design. It found that a ResNet-18 reached 70–76% but attended to unphysical background artifacts, while tree models on color contrast stayed physical but missed thin flakes. A spatial model that ignores physics is therefore not acceptable, and our trust map has to be checked against physical ground truth rather than taken at face value.
Two device papers motivate the stakes, because their usefulness hinges on getting thickness right at a threshold. CuVP2S6 reports ferroelectric synaptic devices built from flakes down to 15 nm. Narrow-bandgap graphene nanoribbons reports a 7 nm length threshold for end-state decoupling.
Research Program
The work begins with a physics audit. For each substrate in the MaskTerial datasets (graphene, WSe₂, hBN, WS₂, MoSe₂ on varying SiO₂ thicknesses), we will use transfer-matrix calculations to compute the color of an n-layer flake under a polymer film of thickness t. The film will have a PMMA- or adhesive-like refractive index near 1.5, with t swept from 0 to a few hundred nanometers. We then find the residue thicknesses at which that color falls within measured camera noise of a clean (n+1)-layer or (n−1)-layer flake. This produces an aliasing map. If aliasing is negligible everywhere, the problem reduces to simple anomaly detection on color, and we would redirect the project toward that cheaper solution.
Next we build an occlusion renderer on top of MaskTerial's pipeline. It composes three kinds of obstruction, each rendered physically rather than pasted on. Residue films get spatially smooth random thickness fields and blob-like outlines. Overlapping flakes are treated as true multilayer stacks, whose color the transfer-matrix method computes correctly. Opaque bulk shards cast shadows and add scattering. Every rendered pixel carries its true flake thickness and a binary visibility label marking whether its color is an unobstructed reading of that thickness.
The model is a segmentation backbone of the MaskTerial type with three per-pixel outputs: a thickness class, a visibility score, and an instance mask. Certification is a fixed rule applied after the network. A region is certified as layer count n only if it is a connected component whose pixels are predicted as n with visibility above τ, and whose area exceeds the device footprint. Here τ is the visibility threshold, set on validation data to meet a target false-certification rate. Everything else is reported as "unverified" rather than guessed.
Baselines
The naive reference point is a per-pixel Gaussian mixture on substrate-normalized color. It labels each pixel by its nearest thickness cluster and certifies connected components of the target class. This is the standard pre-deep-learning practice, and MaskTerial reports outperforming such baselines by over 10% in AP50 on clean data. It will be fit on the same calibration images every other method sees.
The strongest existing method is MaskTerial itself, run as released, with certification taken from its instance thickness label after rejecting instances whose deterministic uncertainty exceeds a threshold. That threshold is tuned on the same validation split and to the same false-certification target as ours. MaskTerial is the strongest competitor for two reasons: it already has physics-based pretraining, and it has an uncertainty signal that a reviewer will argue should already catch residue. No absolute AP figure for it is given in the file, and for occluded or residue-covered flakes we found no reported result on Lacuna for any method. Beating it means a lower false-certification rate at matched certified area, with no more than a small drop in clean-data AP50.
Two ablations isolate our ingredients, and both use the same backbone, compute and tuning budget as the full model. The first fine-tunes MaskTerial on our occlusion renders with no visibility head, abstaining via its own uncertainty. This tests whether the data alone does the work. The second keeps the visibility head but trains it on residue pasted in as random color patches rather than physically rendered films. This tests whether physical rendering matters.
The fourth comparison is a physically interpretable tree model on quantized color contrast, as in Physically informed machine-learning algorithms, applied per pixel. It represents the reviewer's preference for interpretable contrast rules.
Why not just…?
Why not just classify each pixel's contrast and take a majority vote over the flake? Thickness really is a local optical property, so this is the natural first try. It works only where residue pushes colors outside every thickness cluster. The physics audit measures exactly how often residue instead lands inside another cluster. In aliasing regimes a per-pixel classifier certifies residue-covered monolayer as bilayer, or the reverse, with full confidence. If the audit shows little aliasing, this baseline will match us and we will say so.
Why not just threshold MaskTerial's existing uncertainty? Deterministic uncertainty flags inputs that lie far from the training density. Aliased residue pixels lie inside the density of a legitimate class, so we expect this to miss them. This is the claim most at risk. If MaskTerial's uncertainty matches our visibility head on the aliased subset of the test set, the visibility head adds nothing.
Why not just augment the training data with occlusions and skip the visibility head? This is the real threat, and the first ablation exists to settle it. If augmentation alone reaches the same false-certification rate, the contribution reduces to the physically rendered occlusion dataset and the abstention protocol. That would still be useful, but it is a smaller claim, and we would report it as such.
Why not just measure every candidate with Raman spectroscopy or atomic force microscopy? Both are reliable, and we use them as ground truth. They are serial point or line scans, which the cited flake-identification work identifies as the throughput bottleneck. The tool's job is to decide which handful of regions deserve those measurements.
Proposed Experiments and Decisive Checks
We will evaluate on three test sets.
The first is clean data: the original MaskTerial datasets, the Masubuchi dataset as used by CLIFF, and the MoSe₂ set from the physically informed study. This checks for regressions. For context, CLIFF reports 56.96% final macro accuracy on Masubuchi, but in a continual-learning setting, so it is reported for reference only and not as a direct comparison.
The second is a held-out synthetic occlusion set with exact labels. It includes a stratum of pixels that the audit marks as aliased.
The third and most important is a real occluded set. We will select roughly a few hundred microscope fields from the MaskTerial materials that contain visible residue, overlaps or terraces. Each method's certified regions will be checked by Raman mapping or atomic force microscopy, with spot checks of uncertified regions so we can estimate recall.
The primary metric is the false-certification rate at matched total certified area, which we will report as a risk-coverage curve. Secondary metrics are AP50 on clean data, pixel thickness accuracy restricted to visible pixels, and calibration of the visibility score against renderer labels.
Four results would support the proposal: the physics audit finds substantial aliasing for at least the low-contrast materials; the full model beats MaskTerial-with-uncertainty and the augmentation-only ablation on the aliased stratum and on the Raman-verified real set; physically rendered residue beats pasted residue; and clean-data AP50 holds. If the augmentation-only ablation ties on the real set, we will narrow the claim to the dataset and protocol.
As a final device-level check, we will fabricate a small batch of field-effect transistors for one material, on monolayer regions certified by our method and by MaskTerial. This tests whether lower false certification translates into fewer devices with off-target behavior. It ties the vision result to the integration-yield problem that motivates the work.
References
- MaskTerial: a foundation model for automated 2D material flake detection
- CLIFF: Continual Learning for Incremental Flake Features in 2D Material Identification
- Learning to Know Where to See: A Visibility-Aware Approach for Occluded Person Re-identification
- Pose-Guided Visible Part Matching for Occluded Person ReID
- High-Order Information Matters: Learning Relation and Topology for Occluded Person Re-Identification
- MHSA-Net: Multihead Self-Attention Network for Occluded Person Re-Identification
- Physically informed machine-learning algorithms for the identification of two-dimensional atomic crystals
- Intrinsic ferroelectric CuVP2S6 for potential applications in neuromorphic recognition and translation
- Growth Optimization and Device Integration of Narrow‐Bandgap Graphene Nanoribbons