GaN is the one wide-gap semiconductor that already ships in volume — LEDs, lasers, RF, and now power — and every one of those products is limited by defects that do not show up in a wafer map. Dislocations are counted routinely. Point defects, the nonradiative centers that quietly set the internal quantum efficiency of an InGaN quantum well, are not. They are too small to image and too dilute to see in an average spectrum.

Hyperspectral emission maps can see them, if you look at the right statistic. This post takes a published cathodoluminescence dataset of InGaN/GaN quantum wells with a controlled point-defect series and shows what Atlas extracts from it with the same Blocks that run in the Matter42 app — and where a dedicated III-nitride analysis would take it further.

The dataset

Weatherley and co-workers grew five InGaN/GaN single-quantum-well samples that differ in one thing: the number of periods in an InGaN/GaN superlattice underlayer beneath the well, from zero (P0) to twenty-four (P24). The underlayer is known to suppress a nonradiative point defect in the well. Time-resolved PL showed carrier lifetime and internal quantum efficiency rising with underlayer thickness, and by counting dark spots in their CL images the authors put the density of the efficiency-limiting defect at 10¹⁶ cm⁻³ in P0 and 10¹⁴ cm⁻³ in P24. They then mapped the same areas of each sample with hyperspectral cathodoluminescence at 10 K and 170 K, and P24 again at 300 K, and published the raw cubes alongside the paper.

Two honest framing notes before the results.

These are cathodoluminescence maps, not photoluminescence. Atlas preserves them as CL datasets while using the shared emission-map representation needed for wavelength, linewidth, intensity, and spatial-quenching analysis. The excitation volume and carrier density differ from optical PL, which is why the CL workflow keeps the technique label and electron-beam acquisition context explicit.

And the material was declared to Atlas as InGaN, not GaN. The QW emits near 406 nm; bulk GaN's band edge is at 364 nm. Given material=GaN, Characterize CL would latch onto the weak GaN barrier shoulder and miss the well entirely. Telling the Block which layer you care about is part of the measurement.

One sample, one pass

Characterize CL fits the primary emission line on the map mean, then per pixel, and reports the shift against the material's reference, the linewidth, and map statistics for each.

Two things to notice. The linewidth-based defect score is low and the Block says so; by the criteria that flag a disordered TMD or a defect-banded III-V film, P0 is clean. But the intensity map is not uniform. It is peppered with dark spots, and the standard deviation of integrated intensity across 16,384 pixels is 17.6% of the mean. That is the point-defect signature the paper is built on: each nonradiative center creates a dark disc of roughly a carrier diffusion length around it.

Here is the sample at the other end of the series.

Same Block, same parameters, same map size. The intensity spread has fallen from 17.6% to 4.7% of the mean, and the dark spots are almost gone. The linewidth also narrowed, from 66 to 58 meV — a smaller effect, and one that alloy disorder in the well competes with, but in the same direction.

The −54 meV shift on P24 versus −7 meV on P0 is not a defect signal. The five samples are five different wafers with slightly different indium content and well width, so the absolute emission energy moves around by tens of meV between them. Atlas reports it as a shift against the InGaN reference and offers the standard three explanations; it does not pretend to know which one applies.

The whole series

Running the same Block on all five 10 K maps and plotting the one statistic that tracks the paper's independent defect-density measurement:

The intensity spread falls monotonically with underlayer thickness. The 5th-percentile pixel — a rough "how dark are the dark spots" number — climbs from 73% of the mean to 92%. Neither the emission energy nor the linewidth does anything so orderly, which is what you expect: those are dominated by the alloy, and the alloy was not what changed between samples.

This is not the paper's analysis. The authors fitted every pixel, located individual dark spots, classified them by their temperature behavior, and counted them into a volume density. Atlas has not done that here, and we have not claimed a defect density from these maps. What Atlas has done is take a raw cube from each sample, fit the well emission per pixel, and return a spatial-uniformity statistic that orders the five samples exactly as the independent lifetime measurement does — without being told which sample was which.

For a nitride growth team, that is the useful version: a number that comes out of a routine map, with no fitting decisions made by hand, that moves in the right direction when the point-defect density moves.

The same sample, warm

The 300 K series on P24 was acquired at a different probe current and integration time, so intensities are not comparable to the 10 K maps. Line shape is.

At 300 K, Characterize CL flags the 104 meV linewidth as "well above the ~46 meV room-temperature thermal floor — disorder, doping, or unresolved multi-peak structure". That is the correct flag for an InGaN well, where alloy fluctuations broaden the line well beyond kT even in good material, and it is why the room-temperature linewidth is a weak quality metric for this system while the 10 K intensity uniformity is a strong one. Knowing which observable to trust at which temperature is most of the job.

What a full GaN workflow adds

This dataset exercises one Block because it contains one kind of measurement: a quantum-well line in a 350–480 nm window. Production GaN characterization usually has two more, and Atlas has dedicated III-nitride Blocks for both.

GaN PL defect bands takes a wider PL spectrum — near-band-edge through the yellow band — applies the Jacobian correction into energy space, fits the Reshchikov band taxonomy (UVL, BL, GL, YL, RL), and returns the window-integrated YL/NBE ratio that is the standard carbon-contamination and quality metric. It also warns when the excitation wavelength cannot reach the band edge, which makes the ratio meaningless.

GaN Raman stress & doping fits the E₂(high) and A₁(LO) phonons, converts the E₂(high) shift to biaxial stress in GPa with a stated coefficient, and inverts the residual A₁(LO) shift through the LO-phonon–plasmon coupling relation to a free-carrier concentration — on maps, as stress, log₁₀ n, and E₂(high) linewidth images. It states its own systematics: the stress coefficient's factor-of-two literature spread, the n-type assumption, the 10¹⁷ cm⁻³ floor, and the 4H-SiC substrate line that sits inside the A₁(LO) window.

Neither ran here because the data does not contain what they need, and we would rather show one real result than three illustrative ones. If you have GaN-on-Si or GaN-on-SiC Raman maps and yellow-band PL from the same wafers, those are the Blocks that turn them into stress, carrier density, and a carbon proxy.

What this gives a nitride team

From six raw hyperspectral cubes, without hand-fitting:

• per-pixel quantum-well emission energy, linewidth, and intensity maps for each sample, with the material reference chosen for the layer of interest,
• a spatial-uniformity statistic that orders the five-sample point-defect series correctly against the paper's independent lifetime and defect-count measurements,
• a dark-spot depth proxy (5th-percentile / mean intensity) that improves from 0.73 to 0.92 across the same series,
• a temperature comparison that measures the 58 → 104 meV broadening and flags why room-temperature linewidth is the wrong quality metric for an InGaN well,
• and an explicit statement of what was not done — no individual defect counting, no volume density — so the numbers above are not mistaken for the paper's.

The Blocks used are documented at Characterize CL, GaN PL defect bands, and GaN Raman stress & doping. Every figure above was generated from the published cubes through the same Atlas analysis functions the Blocks run.

If you grow or process III-nitrides and have emission maps, Raman maps, or yellow-band spectra you would like read this way, request access or get in touch.