A Nature Communications paper says plasma etching can carve monolayer MoS₂ into nanoribbons only tens of nanometers wide without seriously damaging the material.

We got the raw spectra behind that claim.

Instead of reproducing the published figures by hand, we loaded nearly 3 GB of Raman and photoluminescence maps into Atlas and started asking questions in natural language: Did lithography broaden the Raman modes? Did the optical response change globally? Can we recover the reported edge blueshift? Could strain, defects, or carrier density explain it?

One of the first PL answers looked dramatic.

It was also wrong.

That turned out to be one of the most useful parts of the analysis.

Lan et al. recently reported that narrowing monolayer MoS₂ transistors from 540 nm to about 35 nm increased median on-current by roughly 42% and improved subthreshold swing by about 16%. Their paper, published in Nature Communications as "Scaling two-dimensional semiconductor nanoribbons for high-performance electronics," attributes the device improvement to a combination of minimal edge-induced disorder, stronger gate electrostatics, and more efficient side-contact injection.

A basic materials question sits underneath that result: can a Cl₂/O₂ plasma define a 30–70 nm MoS₂ ribbon without introducing enough damage to undermine the device?

The paper makes three important optical observations:

1. The characteristic Raman modes do not measurably broaden after patterning.
2. The A-exciton energy increases near an MoS₂ edge, with the main text reporting an edge-versus-bulk blueshift of about 5 meV.
3. Natural and lithographically etched edges show similar spatial PL behavior.

The authors interpret the PL shift as being consistent with reduced electron density near the edge and propose oxygen passivation of sulfur vacancies as one possible microscopic explanation.

We wanted to see what the raw maps themselves would support.

Six maps, nearly three gigabytes

Collaborators shared the raw hyperspectral maps behind those measurements: two 60×60-pixel Raman maps at 532 nm, three PL maps crossing film edges, and one PL map of the finished ribbon-array device. Together, the text exports occupy about 2.9 GB.

The PL data came from a WITec alpha300R and the Raman data from a Thermo Scientific DXR3xi. Each hyperspectral export contains roughly 1,600 spectral channels per spatial pixel.

Atlas parsed the measurements into the same dataset model while retaining their spatial grids, excitation wavelengths, quality information, and links back to the original files.

We then investigated the project through the Atlas agent in natural language. Underneath those questions, the agent called reusable analysis Blocks: explicit analysis units with defined inputs, parameters, outputs, warnings, and provenance.

The Block outputs embedded below are captured results from this analysis. The Blocks themselves can be run again on another dataset with the same parameters, inspected, and forked or customized when a different analysis is needed.

We did not rely on pixel-level Raman × PL fusion across the full dataset because robust co-registration was not available for all maps. The useful comparisons here are condition-level and geometry-aware: patterned versus unpatterned Raman, edge versus interior PL, and whether independent observables tell a consistent physical story.

First test: did lithography broaden the Raman peaks?

This is the cleanest test of the paper's optical damage claim.

Atlas fit the characteristic E′ and A₁′ Raman modes in the patterned and unpatterned maps using the same Fit peaks Block.

The spectra nearly overlap.

For the captured Block runs, the patterned and unpatterned linewidths differ by no more than 0.11 cm⁻¹:

There is no systematic widening of both modes after lithography.

That part of the paper holds up well. At the spatial scale resolved by the Raman measurement, the patterned region shows no evidence for a large etch-induced increase in disorder.

There is an important limit to that statement. The Raman spot was approximately 1 µm across, much larger than a 30–70 nm ribbon edge. The measurement can rule out broad, map-scale damage much more confidently than it can rule out a very narrow damaged zone localized to the physical edge.

That distinction matters throughout the analysis: "no detectable Raman broadening" is not the same claim as "the atomic edge is defect-free."

The first PL analysis gave the wrong answer

The PL result is more subtle.

The paper used K-means clustering to separate bulk-like and edge-like spectral populations and reported an approximately 5 meV blueshift of the edge population.

So an obvious first attempt was to ask Atlas to divide the raw PL maps into two spectral populations.

The clustering looked convincing. It produced distinct spatial regions and large differences in apparent peak energy.

Too large.

For the well-covered 532 nm maps, the apparent cluster separations were roughly:

Those values are an order of magnitude larger than the main-text result.

The spatial maps showed why.

K-means had found the largest spectral contrast in the map, but that contrast was not necessarily the physical contrast we had asked about. One population contained genuine MoS₂ excitonic PL near 1.81 eV. The other was dominated by weak, non-emitting, or low-SNR pixels where a meaningful exciton peak could not be reliably fit.

A peak finder can still return a number from a nearly flat spectrum. That does not make the number a physical exciton energy.

Calling those clusters "edge" and "bulk" would therefore turn a mathematically clean segmentation into a physically wrong conclusion.

This is an important distinction for automated analysis:

unsupervised clustering finds dominant variation; it does not know which variation answers the scientific question.

K-means itself is not the problem. The result depends on what features dominate the clustering and which pixels are included. Here, signal-versus-no-signal contrast overwhelmed the much smaller edge-versus-interior spectral effect we actually wanted to measure.

The failure also explains why simply averaging an entire PL map is not enough.

A whole-map mean mixes a narrow edge region with a much larger interior population. In the best-matched Litho 532 nm and NoLitho 532 nm maps, the global mean exciton energies differ by only about 0.3 meV and the fitted linewidths by about 0.1 meV.

That is useful evidence that lithography does not grossly alter the overall PL response.

But it cannot test a localized edge effect.

The scientific question itself is spatial:

What happens to the PL peak as we approach the film edge?

That required a different analysis.

A better test: profile the spectrum by distance from the film edge

We replaced the unconstrained cluster labels with a geometry-aware analysis using Profile a metric vs distance from film edge.

The Block bins a per-pixel spectral metric by its distance from the nearest film edge and reports the median trend, the interior bulk reference, the edge value, and a fitted exponential decay length. It runs at the map's native pixel pitch, and peak-center measurements use sub-pixel estimation by default.

For these runs, we used the Block's spectral-signal edge option rather than its default parse-time quality mask. This option uses a spectral-signal SNR gate to define the film boundary before profiling the PL peak center versus distance from the edge.

For PL peak-center profiles, Atlas displays the result as an energy shift relative to the Block's interior bulk reference, in meV.

That distinction is important. The reported edge value is not an absolute exciton energy. It is the Block's reported edge value for a profile whose PL peak center is expressed relative to the interior reference.

The captured runs report:

• Lithographic edge, 532 nm: edge value +12.9 meV
• Natural edge, 532 nm: edge value +11.4 meV
• Lithographic edge, 488 nm: edge value +10.6 meV

In all three runs, the PL peak-center profile shifts toward higher energy at the edge relative to the interior reference.

The two 532 nm edge values are also strikingly similar in magnitude.

This needs to be compared carefully with the paper.

The approximately 5 meV value in the main text comes from a K-means-separated edge-versus-bulk population comparison. Figure 5 in the paper's supplementary information instead examines how the fitted PL peak changes as the measurement approaches natural and lithographically defined edges.

Those are different summaries of spatially varying data.

A broader edge-like population can mix pixels at several distances from the physical boundary, whereas a distance profile preserves that spatial dependence.

The quoted <5 meV value for the instrument is an absolute spectral-accuracy specification. By itself, it does not determine the uncertainty of an edge-to-bulk relative shift measured within a map. We therefore do not assign formal statistical significance to these edge values without a dedicated uncertainty analysis.

The strongest conclusion is not that Atlas reproduces the main-text 5 meV number digit for digit.

It is that an independent, geometry-aware analysis of the raw maps recovers the same qualitative spatial structure:

• the PL peak shifts toward higher energy at the edge,
• the change is localized to the edge region rather than appearing as a large global shift,
• natural and lithographically defined edges show similar behavior,
• and the two 532 nm runs return similar edge values.

The paper already contains a distance-dependent PL analysis in Figure 5 of the supplementary information, so the novelty here is not simply the idea of plotting PL versus distance from an edge.

The addition is that Atlas turns it into a reusable analysis operation on the raw hyperspectral map: the film boundary can be derived from the data, a per-pixel spectral metric can be profiled against distance from that edge, and the same Block can then be applied to additional observables.

That makes it possible to ask another question without building a new one-off analysis.

A second check on electron depletion

If the blueshift reflects lower electron density, the lower-energy trion-related contribution should also become less prominent near the edge.

We therefore used the same Profile a metric vs distance from film edge Block with metric='fraction'.

For this run, the configured spectral windows measure the low-energy contribution as approximately:

I(1.775–1.815 eV) / [trion + exciton band]

Using the same edge-analysis settings lets the low-energy band share be examined against the same physical variable: distance from the film edge.

Relative to the Block's interior reference, the captured runs report an edge change of:

• Lithographic edge: −1.0 percentage point
• Natural edge: −5.6 percentage points

The low-energy band's fractional contribution therefore falls near both edges.

That trend is consistent with electron depletion.

It is not a chemical identification, and the absolute spectral assignment deserves some care. The low-energy band can mix trion and bound-exciton emission, so its spatial trend is more reliable than assigning the entire band to one microscopic species.

When a fitted X⁻ contribution is preferred over a fast band fraction, the same Block also accepts metric='trion_weight', which runs the same per-pixel exciton/trion decomposer used to fit the neutral-exciton/trion split directly.

The useful point is that the edge interpretation no longer rests on only one PL observable.

The PL peak energy and the lower-energy spectral fraction change in the same physical region and point in a consistent direction.

Could strain explain the PL edge behavior?

Atlas also examined the Raman maps for an alternative explanation.

The patterned Raman analysis shows little correlation along the fitted strain axis, with r = −0.10 and an estimated strain spread of about 0.066%.

That makes a large, map-scale strain change an unlikely explanation for the PL trend.

But the Raman measurement cannot rule out all strain-related edge physics.

Its approximately 1 µm optical spot is much larger than the nanoscale edge itself, and the Raman and PL measurements are not being treated here as a fully co-registered pixel-level dataset.

The defensible conclusion is therefore narrower:

the Raman maps show no strong large-area strain signature introduced by patterning, but they cannot exclude localized strain within an optically unresolved edge region.

What about defects?

Atlas also ran an independent LA(M)-based defect proxy on the two Raman maps using Estimate defect density.

The captured results:

• Patterned: 0.44% LA(M)-based estimate, 3.7 nm estimated defect correlation length
• Unpatterned: 0.36% LA(M)-based estimate, 3.9 nm estimated defect correlation length

The two maps therefore sit on a similar calibration-derived defect scale, with no evidence here for a large etch-induced increase.

These values require a stronger caveat than a fitted Raman linewidth.

They come from a literature-calibrated LA(M) relation, so the percentages should not be read as direct atomic defect counts.

The corresponding L_D values are estimated defect-correlation-length quantities, not microscopy measurements of literal average spacing between individually observed defects.

Taken together with the nearly unchanged Raman linewidths, the defect proxy gives another indication that patterning did not produce a large increase in Raman-visible disorder.

The paper's oxygen-passivation explanation remains a plausible microscopic interpretation, but these optical measurements do not directly establish oxygen incorporation. A chemically specific measurement would be needed to make that assignment directly.

What survived the stress test?

The raw-data analysis gives a more useful answer than simply "replicated" or "not replicated."

The most useful lesson came from the analysis that initially gave the wrong physical answer.

K-means produced a clean numerical result that did not correspond to the question we wanted to answer. Inspecting the spectra and spatial regions exposed the problem. Recasting the analysis around distance from the film edge produced a much more meaningful comparison.

That is closer to how scientific analysis actually works than a single push-button "replication."

These analyses are reusable

None of the quantitative analyses in this post need to remain a one-off notebook.

They are implemented as Atlas Blocks: reusable analysis units with explicit inputs, parameters, outputs, quality information, and provenance.

You can run these same Blocks yourself on your own Raman or PL datasets, inspect how the result was produced, reuse the same workflow, and fork or customize the analysis when different assumptions are needed.

The main Blocks used here are:

• Fit peaks
• Profile a metric vs distance from film edge
• Estimate defect density
• Decompose strain vs doping

The cards embedded in this article are captured snapshots of the analysis used for this case study.

Inside Atlas, the corresponding Blocks remain executable analysis tools tied to the underlying datasets rather than static figures detached from their source.

Raw data should outlive the figure

The Raman result is reassuring: the raw maps independently support the paper's conclusion that lithography does not introduce a large, Raman-resolvable increase in disorder.

The PL story is more interesting.

A whole-map comparison says the patterned and untreated material look globally similar.

Generic whole-map K-means clustering gives a dramatic but physically misleading answer.

Profiling the spectral response by distance from the film edge then recovers a localized positive PL-energy shift at both natural and lithographically defined edges. Applying the same type of edge-profile analysis to the lower-energy PL fraction adds a second trend consistent with reduced electron density.

The result is not simply a reproduction of a published figure.

It is a connected chain of evidence:

unchanged Raman linewidths, globally similar PL, a localized edge-associated energy shift, a matching low-energy PL trend, no strong map-scale strain signature, and similar calibration-based defect proxies before and after patterning.

Just as importantly, each conclusion stays tied to what the measurement can actually resolve.

That is what becomes possible when the raw data behind a publication stays usable.

A published figure can be checked. An analysis can be challenged and refined. A claim can be narrowed to what the measurement actually supports. And the same dataset can answer questions beyond the original figure.