MoS₂ is the 2D semiconductor closest to a fab. It has the most deposition routes — exfoliation, CVD, MOCVD, ALD, sputtering, sulfurization of Mo — and therefore the widest spread in what "MoS₂ film" actually means on the wafer. Two films with the same thickness and the same XRD phase can differ by three orders of magnitude in transistor on/off ratio.

Raman spectroscopy is the fastest way to tell them apart, and the difference is written into the spectrum in ways a quick peak-position readout misses. This post takes four real MoS₂ films from a deposition study — ALD and sputtered, each with and without a post-deposition anneal — and shows what Atlas extracts from a 25-pixel Raman map of each, using the same Blocks that run in the Matter42 app. The device data from the same study is then used to check whether the spectroscopy predicted anything.

We have written about MoS₂ before: a single AFM scan pushed as far as it goes, and a stress test of a published nanoribbon result on 3 GB of Raman and PL maps. This post is the deposition-process view.

Four films, one instrument

A collaborating device lab grew MoS₂ by two routes and measured both as-deposited and after annealing. Each condition was mapped on a Horiba LabRAM HR Evolution at 532 nm with an 1800 gr/mm grating; the study also fabricated back-gated FETs on every film.

Atlas parsed all four LabSpec exports into the same hyperspectral dataset model and identified the material as MoS₂ from the Raman modes. A 5 × 5 map is small — this is a spot check, not a wafer map — but it is what a process engineer actually has at the end of a deposition run, and 25 spectra is enough for the statistics below to mean something.

You can see the answer in the mean spectra. The rest of the post is about making it a number, and about which numbers Atlas will and will not stand behind.

Linewidths: the primary quality readout

Fit peaks fits a pseudo-Voigt to each mode in every pixel and reports the map-mean spectrum fit, the per-pixel statistics, and the significance of each peak.

Across all four films, same windows, same Block:

The ALD films sit at the textbook few-layer positions with linewidths a factor of three narrower than the sputtered films. Annealing tightens the ALD E₂g by about 0.9 cm⁻¹ and does very little for the sputtered films' width, though it does pull the sputtered E₂g 2.4 cm⁻¹ back toward its proper position.

One number in that table deserves a caveat, and it is a useful one. The sputtered E₂g width depends on the fit window: 16 cm⁻¹ with a tight 374–392 window, 23 cm⁻¹ with a generous 365–395 one. The ALD films give 5.7–6.0 cm⁻¹ for any of them. The window sensitivity is itself a disorder signature — it means there is a shoulder under E₂g that a single symmetric profile cannot separate — and Atlas exposes the window as a parameter rather than burying the choice.

Mode separation: thickness, or something else?

The A₁g − E₂g separation is the standard thickness proxy for MoS₂: about 19 cm⁻¹ for a monolayer, growing to ~25 cm⁻¹ in bulk. Atlas applies that calibration and, on maps, also checks whether the in-plane mode position varies enough to be confused with a layer step.

The ALD films read as two to three layers, consistently. The sputtered films read as "bulk" — but with a separation of 29–33 cm⁻¹ that exceeds the bulk value by a wide margin, which is not a thickness at all. It is a disordered film whose E₂g has been pushed down by 4–7 cm⁻¹. Atlas does not pretend otherwise: the layer call comes with a warning that the in-plane mode position spans a layer step's worth of variation across 25 pixels, and points to the strain/doping decomposition as the cross-check.

This is the general lesson for anyone using Δω as a thickness monitor on deposited (rather than exfoliated) MoS₂. The calibration assumes a well-ordered crystal. Off that assumption, the number still comes out — and it is wrong in a way that looks like "thick".

Where the disorder lives: strain and doping spread

Decompose strain vs doping takes the per-pixel E₂g and A₁g shifts and projects them onto the strain and carrier-density axes using published response coefficients (Lloyd et al. for biaxial strain, Chakraborty et al. for doping — both measured on MoS₂, so no transferred constants here).

Within a 25-pixel map the ALD films vary by three hundredths of a percent in strain. The as-deposited sputtered film varies by seven times that, and its apparent carrier density by eight times. Annealing the sputtered film halves the strain spread and cuts the doping spread by a factor of four — the clearest quantitative statement in this dataset of what the anneal actually did, and one that neither the peak positions nor the linewidths alone made.

Read these as spreads, not absolutes. The decomposition is relative to the map's own reference point, and with the E₂g and A₁g shifts essentially uncorrelated in every film (|r| < 0.2), Atlas notes that the variation is dominated by a single axis or by noise. For the ALD films, "noise" is the right reading: they are uniform to the instrument's limit.

Defect density: two estimators, and when to trust neither

Estimate defect density runs two independent inversions on MoS₂: the E₂g linewidth against a simulated defect-supercell calibration, and the disorder-activated LA(M) band's height ratio against the E₂g mode (Mignuzzi et al., PRB 91, 195411 (2015)), which returns a defect correlation length L_D.

Two honest things happen in this table.

For the sputtered films, the E₂g linewidth falls outside every calibration curve Atlas has. The Block reports that the density it would return is an extrapolation and flags it. We have not quoted the number. A tool that returned "10% sulfur vacancies" for an 18 cm⁻¹ linewidth would be producing a figure, not a measurement.

For the ALD films, both estimators return values, and they disagree by a factor of three to four. Atlas flags that too. The linewidth calibration is built from point-defect supercells; the LA(M) ratio is empirical. When they diverge, either the dominant disorder is not the point-defect type either calibration assumes, or there is another broadening contribution — and for a 2–3 layer ALD film, interlayer stacking disorder is a good candidate. The LA(M) estimate sits on a direct MoS₂ calibration at this wavelength and is the one worth carrying forward for the ALD pair: 0.38% annealed versus 0.53% as-deposited, a 30% reduction in disorder-activated scattering from the anneal. It is not a route discriminator — the as-deposited ALD and sputtered films land at the same ratio — because a Raman-active LA(M) band needs a lattice coherent enough to define an M point, and the sputtered films are partly past that.

Did the spectroscopy predict the transistor?

The same study fabricated back-gated FETs on all four films. The medians from those measurements, taken from the collaborators' own analysis and not from Atlas:

The Raman linewidth sorted the two deposition routes correctly before a single contact was patterned. That is the practical value: a five-minute map that tells you whether to bother with the lithography.

It also shows the limit. Within the ALD pair, the anneal narrowed E₂g and reduced the LA(M) disorder ratio, but the on/off ratios are indistinguishable. Within the sputtered pair, the annealed film looked slightly better in every Raman metric and was worse as a transistor — with only two devices measured, that may be statistics rather than physics. Raman sees the lattice. The device also sees contacts, the dielectric interface, and trap states that phonons do not report on. A film that fails Raman will fail as a device; a film that passes Raman has earned the right to be tested, not a guarantee.

What this gives an MoS₂ team

From four 25-pixel maps:

• fitted mode positions and linewidths on a common window, with the sputtered E₂g width flagged as window-dependent because of the shoulder underneath it,
• a layer count for the ALD films (2–3 L) and an explicit warning that the "bulk" call on the sputtered films is disorder masquerading as thickness,
• strain and doping spreads that quantify what the anneal did to the sputtered film — strain spread halved, doping spread cut by four — and confirm the ALD films are uniform to the noise floor,
• an LA(M)-based disorder estimate that resolves the anneal within each route, alongside a linewidth-based one that Atlas itself declined to trust for two of the films,
• and a structure–property check against real devices that says the map predicts the deposition route, not the anneal.

Atlas did not identify the dominant defect species here. Classify defect type ran on every film; the best-match scores were 0.40–0.60 and every top candidate's estimated density fell outside its calibration range, so Atlas marked the rankings qualitative and we have not quoted them. On 25 pixels at a single excitation wavelength, that is the correct answer. A 633 nm map of the same film — where the disorder-activated modes are resonantly enhanced — is the measurement that would sharpen it, and Atlas will read the wavelength from the file and adjust its resonance handling accordingly.

The Blocks used here are documented at Fit peaks, Decompose strain vs doping, Estimate defect density, and Classify defect type. Every figure above was generated from the four source files through the same Atlas analysis functions the Blocks run.

If you deposit MoS₂ — by any route — and want a quality readout you can compare across runs and against your devices, request access or get in touch.