Labs that build devices from two-dimensional crystals, meaning sheets one to a few atoms thick such as graphene, tungsten diselenide or black phosphorus, now use neural networks to scan microscope images and find usable flakes. The strongest of these networks learn almost entirely from simulated images. The simulator computes how light reflects off a thin stack of layers, so it can produce unlimited labeled examples. That simulator assumes every flake is pristine. Many of these materials are not: in air they grow oxide skins, lose layers from the edges inward, and turn into mixed phases. An oxidized bilayer can reflect light much like a monolayer. A detector trained on pristine physics may then report it as the prized target, and nobody finds out until a device fails.
We propose to put oxidation into the simulator itself. Each synthetic flake becomes a stack of remaining pristine layers under an oxide layer whose thickness varies in space, and the network learns to report both the original layer count and how degraded the flake is. The claim is tested where the simulation literature usually stops: on real time-lapse aging series, in which the same flakes are imaged fresh inside a glovebox and then repeatedly as they age in air. Because the fresh image fixes each flake's identity and thickness, the aging series supplies real-world labels without anyone annotating an oxidized image by hand.
Core Claim
A physics-based synthetic generator that models oxide overlayers and edge-inward oxidation fronts will let a flake detector keep telling target flakes apart from look-alikes as samples age. That means rejecting degraded thicker flakes and oxidation residue that mimic the target's contrast. It should do this better than the same detector trained on pristine simulation, and better than one given the same small budget of real aged images for fine-tuning. The second comparison carries the real burden, and the design is built around it.
The idea draws on three bodies of work. MaskTerial supplies the method. It generates 300,000 synthetic images from 35,000 real flake outlines using the transfer-matrix method, a standard calculation of thin-film interference colors. It pre-trains an instance segmentation network on them and closes the remaining gap with 2 to 5 real images per class plus an uncertainty head. Its optics assume uniform, unoxidized layers, and it never tests on aged samples; we keep its pipeline and change the physics.
Several papers motivate the gap by documenting that oxidation produces intermediate phases rather than clean failure. These include secondary SnS₂ forming on SnS during exfoliation (Xu et al.), TiO₂ islands on MXenes (Cai and Kim), rapid photo-oxidation of phosphorene (Zhang et al.), and MoOₓ at exposed MoS₂ edges (Jain et al.).
A classical detection pipeline constrains the design. The open-source GMM platform shows that contrast alone produces false positives from tape residue, and it needs shape filtering. This tells us background look-alikes are the dominant failure, so a generator must model backgrounds as well as flakes.
From rare-event physics we borrow the reporting convention: background rejection at a fixed signal acceptance. Examples are the 90%/90% operating point of Holl et al. and the surface-alpha rejection at 0.2% bulk loss in Arnquist et al.. A flake survey faces the same asymmetry. Missing a monolayer costs little, but a false monolayer wastes a fabrication run.
Research Program
The oxidation-aware generator. We extend MaskTerial's transfer-matrix renderer. Each synthetic flake is described by a stack that runs from the substrate, through the oxide, through the $n$ remaining pristine layers, to a top oxide layer of thickness $d_{ox}(x,y)$. That thickness field is drawn from a small family of shapes. The family covers a uniform skin, an edge-inward front whose width is set by a diffusion-length parameter, and scattered blister-like spots.
The optical constants of the oxides (phosphorus oxides, tungsten oxides) come from the literature and are randomized within stated uncertainty bands. The research file gives no values, so fitting them is itself a task, calibrated on a handful of real aged flakes. The generator also renders non-flake residue as oxide-only patches with no pristine layers, so the network sees background look-alikes during training.
The model. We keep MaskTerial's Mask2Former backbone and add a second output head for each detected flake. It regresses the fraction of the flake area that is oxidized and the mean oxide thickness, alongside the usual layer-count class. Flakes above a user-set degradation threshold are reported as "degraded" rather than assigned a thickness class. This is how degraded flakes are rejected as background.
The real-world aging benchmark. Flakes of black phosphorus, which degrades in hours, and WSe₂, which degrades slowly, are exfoliated onto SiO₂/Si in a glovebox. They are imaged at time zero and then imaged in ambient air on a schedule running from minutes to days. At time zero, flake outlines and thickness are labeled through the existing pipeline. Thickness is spot-checked by atomic force microscopy or Raman spectroscopy on a subset. The aged frames are registered to the time-zero frame, so every aged flake carries its original label. A subset also receives atomic force microscopy at late times to measure the true oxidation state.
Graphene and hBN, which are air-stable, serve as controls. Their performance should not change with exposure time; if it does, the degradation measured in the reactive materials is confounded by something other than oxidation, such as drifting illumination.
Baselines
The naive reference is MaskTerial as published. It is pre-trained on pristine synthetic images and fine-tuned with its standard few-shot budget on fresh, time-zero images of each material, then run unchanged across the aging series. It shows how much the pristine assumption costs as a function of exposure time. MaskTerial reports beating the GMM baseline by more than 10% AP50 across materials and by up to 40% for hBN. Lacuna has no reported result for any detector on oxidized or aged flakes, so this degradation curve is itself a first measurement.
The classical reference is the GMM plus logistic-classifier platform, which reports AR50 of 63% to 100%. It runs on the same images with its shape filters tuned on time-zero data. It tests whether shape filtering alone already rejects degraded flakes.
The strongest competitor is MaskTerial fine-tuned on real aged images. It receives the same number of labeled real images our method uses for calibration, drawn from the aging series with timestamps disjoint from the test frames. It is the most direct way to "just show the model oxidized flakes," and MaskTerial's own finding that performance saturates at a few images per class makes it a serious threat. For fairness, both arms get the same backbone, the same synthetic pre-training budget in GPU-hours and the same hyperparameter search budget. The only differences are the physics of the synthetic data and the source of the real fine-tuning images.
Beating this baseline means two things. The first is a higher rate of false-target rejection at 90% target acceptance on late-exposure frames. The second is holding that advantage on exposure times and oxidation patterns absent from the fine-tuning images.
Two ablations isolate the key ingredient. One trains on a generator with a uniform oxide skin only, removing spatial fronts. The other trains on pristine physics with aggressive, non-physical color and contrast jitter. The jitter ablation is matched in how widely it spreads the contrast distribution, which separates "physically correct oxidation" from "more diverse contrast."
Why not just…?
Why not just fine-tune MaskTerial on a few aged images? This is the strongest alternative, and it may work at the exposure times it sees. The concern is that oxidation is a continuum. A few images per class cover a few points on the degradation trajectory, while a physics generator covers the whole family. The decisive test holds out exposure times: fine-tune on early and middle frames and test on late frames, and the reverse. If fine-tuning matches our method on held-out times, the physics adds nothing for detection. We would then scale the claim back to the oxidation-state regression head, which fine-tuning on class labels cannot provide.
Why not just use MaskTerial's uncertainty head to flag oxidized flakes as unfamiliar? An uncertainty score may flag heavily degraded flakes. However, a partially oxidized bilayer whose contrast falls within the monolayer's range is by construction in distribution for a pristine model, so it should receive a confident wrong label. We measure this directly by computing the uncertainty baseline's rejection-at-acceptance on the aging series. If uncertainty alone rejects the false monolayers, the proposal reduces to recommending that existing tool.
Why not just image everything in the glovebox, or encapsulate immediately? Samples are routinely handled in air between steps. Jain et al. show that even hBN-encapsulated MoS₂ oxidizes at etched edges. Liu et al. report surface degradation of Na₃Bi inside vacuum. Isolation lowers exposure but does not guarantee pristine samples, and inspection after transfer is where a degradation-aware detector pays off.
Why not just augment pristine simulations with random contrast jitter? This is the jitter ablation above. If it matches the oxidation-aware generator, the gain comes from diversity, not physics, and we will report that plainly.
Proposed Experiments
The primary experiment plots, for black phosphorus and WSe₂, false-target rejection at 90% target acceptance and AP50 against exposure time for all baselines and ablations, with graphene and hBN as flat controls. Success means the oxidation-aware model's curve stays flat, or declines more slowly, where the pristine model's curve falls. It must also match or beat the aged-image fine-tuning baseline on held-out exposure times.
The second experiment tests the oxidation head. We compare predicted oxidized-area fraction and oxide thickness against atomic force microscopy on the late-time subset, and report calibration error. No baseline offers this output.
The third is a transfer test. Generator parameters, including the oxide optical constants, are fitted on WSe₂ and the model is run on black phosphorus with only its literature oxide constants swapped in. This checks whether the oxidation modeling is general or has been overfitted to one material. Chrono:Engine leg evolution is a cautionary case here: designs optimized in one simulated environment performed poorly in others, the same kind of failure this transfer test is meant to catch.
A negative result on the held-out-time comparison would be informative in its own right. It would show that for flake detection, a few real aged images do what a better simulator does. That would be the first measured answer to a question every synthetic-data pipeline in this area currently assumes away.
References
- MaskTerial: a foundation model for automated 2D material flake detection
- Quantum confinement-induced enhanced nonlinearity and carrier lifetime modulation in two-dimensional tin sulfide
- Recent advances in MXene gas sensors: synthesis, composites, and mechanisms
- Advanced optical polarizers based on 2D materials
- One-Dimensional Edge Contacts to a Monolayer Semiconductor
- An open-source robust machine learning platform for real-time detection and classification of 2D material flakes
- Deep learning based pulse shape discrimination for germanium detectors
- $$\alpha $$-event characterization and rejection in point-contact HPGe detectors
- Discovery of a Three-Dimensional Topological Dirac Semimetal, Na 3 Bi
- Towards the targeted environment-specific evolution of robot components