A conventional cathodoluminescence image assigns one brightness value to each electron-beam position. Hyperspectral cathodoluminescence keeps the spectrum instead. Every raster pixel contains an emission spectrum, so the measurement becomes a three-dimensional data cube with two spatial axes and one wavelength or photon-energy axis. That cube can reveal whether a bright region is a band-edge shift, a defect band, an alloy fluctuation, a strain field, or simply more total light—but only after the optical response, scan history, noise, and model assumptions are carried through the analysis.

A hyperspectral CL cube preserves spectral distinctions that a color composite can hide. Let (D(x,y,k)) denote detected counts at spatial position ((x,y)) and spectral channel (k). A panchromatic image collapses the spectral dimension,

while a band map sums only a chosen interval. Peak-energy, linewidth, ratio, and component maps are not raw detector outputs; they are parameters estimated from each spectrum or from a joint cube model. A visually smooth parameter map can therefore reflect regularization, initialization, bounds, or failed fits as much as the material. Raw-count views, fitted residuals, uncertainty maps, and valid-pixel masks belong beside the final visualization.

Instrument calibration turns detector channels into comparable spectra. The wavelength axis is calibrated with traceable or well-characterized emission lines, while a dark acquisition captures detector offset and dark current. A spectral-response correction accounts for wavelength-dependent mirror reflectivity, grating efficiency, slit transmission, window transmission, and detector sensitivity. Flat-field and bad-pixel corrections may be required for array detectors. Cosmic-ray events are sparse and sharp, but blindly smoothing them can also erase real narrow emission; detection should use spatial, spectral, and temporal context and preserve a correction mask.

Photon energy and wavelength obey (E=hc/\lambda), but equal wavelength bins are not equal energy bins. When transforming spectral density, the Jacobian matters:

Peak positions transform directly, but areas, amplitudes, baselines, and line shapes can change under rebinning. A comparison should state whether spectra are displayed per unit wavelength or per unit energy, whether photon counts or radiant power are represented, and whether the response correction was applied before fitting.

The cube is acquired sequentially, so time-dependent behavior becomes spatial structure. A raster may take minutes or hours. Stage drift, scan distortion, charging, carbon deposition, beam-current drift, temperature change, and evolving defect occupation can all create gradients aligned with scan time. Imaging spectrometers may add wavelength-dependent image displacement, sometimes called image shift or keystone, so different spectral channels do not represent exactly the same specimen coordinate. Fiducial secondary-electron images, interleaved references, orthogonal scan directions, frame-based acquisition, and data-driven registration help separate specimen variation from acquisition history.

Pixelwise peak fitting needs a noise model and a failure state. Photon counts often contain approximately Poisson shot noise plus detector read noise, dark noise, and processing correlations. Least squares with uniform weights overemphasizes bright regions differently from a Poisson likelihood. A generic count model is

where (M) is the physical spectral model and (B) is background. Bounds and shared parameters can stabilize weak spectra, but they also imprint assumptions spatially. Each parameter map should be accompanied by uncertainty, residual, convergence, and boundary-hit maps. Pixels that lack enough information should be labeled invalid rather than forced to return a plausible peak.

question[Define transition, shift, defect, or mixture question] --> design[Choose field, pixels, spectrum, dwell, and dose budget]
design --> calibrate[Acquire dark, wavelength, response, and beam references]
calibrate --> cube[Acquire frame-based CL cube plus registered electron images]
cube --> qa{Stable beam, spectrum, position, and specimen?}
qa -- no --> correct[Revise grounding, dose, cooling, optics, or registration]
correct --> cube
qa -- yes --> preprocess[Mask events, correct response, register, preserve provenance]
preprocess --> explore[Inspect raw spectra, sums, variance, and noise structure]
explore --> model[Fit physical peaks or test constrained factorization]
model --> stress{Stable across rank, starts, windows, and replicate scans?}
stress -- no --> model
stress -- yes --> correlate[Compare with morphology, EBIC, composition, and strain]
correlate --> report[Publish raw views, maps, residuals, uncertainty, and metadata]

Dimensionality reduction summarizes covariance; it does not discover chemistry by itself. After reshaping the cube into a matrix (X) with spectral channels by spatial pixels, principal-component analysis finds orthogonal directions of decreasing variance. It is valuable for estimating noise structure, detecting drift, compressing data, and identifying candidate model complexity, but negative loadings and orthogonality are mathematical properties rather than emission physics. A component dominated by a spectral derivative may represent peak shift; one aligned with scan direction may represent drift. Physical assignment requires reconstructed spectra, spatial context, controls, and comparison with plausible transitions.

Nonnegative matrix factorization uses a model such as

where columns of (W) are nonnegative spectral factors and rows of (H) are nonnegative spatial weights. Nonnegativity often makes factors easier to visualize, but the decomposition is generally non-unique and can split one shifting peak into several fixed components or merge correlated physical sources. Rank choice, scaling, initialization, regularization, background, and local minima affect the result. Repeated starts, withheld pixels, synthetic-mixture tests, residual inspection, and stability across rank are needed before calling a factor an endmember or defect species.

A peak shift map rarely has only one physical cause. Band-edge energy can respond to alloy composition, strain, temperature, carrier density, electric field, doping, quantum confinement, and calibration drift. A composition inference requires an appropriate bandgap–composition relation and bowing parameter; a strain inference requires deformation potentials and stress state; a temperature inference requires a material-specific bandgap model. Hyperspectral CL can reveal correlations and boundaries exceptionally well, but separating causes usually requires EDS or EELS for composition, Raman or diffraction for strain, current-series controls for injection, and temperature or bias controls for fields.

Peak centroids offer a model-light summary,

provided background is removed and a physically meaningful window is chosen. The centroid of two overlapping bands can move even when neither band shifts; only their relative amplitudes changed. Likewise, a broadened peak may represent disorder, unresolved components, a temperature gradient, or instrument focus. Window-sensitivity and alternative-model tests should accompany centroid and linewidth interpretations.

Dose and sampling create a three-way trade among spatial, spectral, and statistical resolution. Smaller pixels, narrower spectral bins, and shorter dwell do not independently improve information. Oversampling the excitation volume raises data size and dose without adding spatial bandwidth. Narrow bins divide photons among more channels and can destabilize fits. Longer dwell improves counts but increases drift and beam-induced change. A pilot cube should measure count rate, feature scale, drift, and dose response; binning can then be chosen from the physical resolution and inferential target rather than from the instrument’s maximum settings.

Reproducible analysis preserves the path from raw counts to every map. The archived dataset should include raw cube, dark and calibration acquisitions, beam energy and current, dwell, pixel pitch, scan order, temperature, optical geometry, grating, slit, detector settings, response curve, correction masks, and software versions. Analysis code should record crop, binning, baseline, spectral domain, model, constraints, starting values, weighting, and rejection rules. Saving only rendered false-color maps prevents later checks for saturation, drift, alternate backgrounds, or overfitting.

Correlative registration turns spectral covariance into semiconductor evidence. Secondary-electron or STEM images locate morphology; EBIC tests charge collection; EDS and EELS constrain composition; Raman, HRXRD, or diffraction constrains strain and phase; time-resolved CL tests dynamics. Registration uncertainty matters when the claimed feature approaches the pixel size or drift correction. A component map that follows a structural interface and agrees with an independent material signal is stronger than an attractive factorization alone.

For semiconductor process learning, the central question is not “how many spectral maps can the cube produce?” It is “which spatially varying emission model survives calibration, drift, dose, noise, rank, fit, and independent-physics tests?” Reading hyperspectral CL through that calibrated-datacube-and-model-identifiability lens turns a massive spectrum image into defensible evidence about optical defects, composition, and strain.