An electron crossing an electron-transparent semiconductor lamella can emerge unchanged, or it can surrender a precisely measurable portion of its energy to the specimen. Those losses arise from collective valence excitations, interband transitions, phonons, and ionization of element-specific core levels. Electron Energy-Loss Spectroscopy (EELS) disperses the transmitted electrons by energy inside a TEM or STEM, connecting nanoscale structure to composition, bonding, dielectric response, and local electronic states in one spectrum.

EELS measures an energy difference, but each spectral region answers a different materials question. If the incident electron has energy (E_0) and reaches the spectrometer with energy (E_t), its loss is

Electrons near Δ(E=0) form the zero-loss peak (ZLP), which records the instrument response together with elastic and very-low-energy scattering. The low-loss region contains plasmons, interband transitions, and other excitations related to valence electrons and dielectric behavior. Farther out, core-loss edges begin when the transferred energy can excite an inner-shell electron into an unoccupied state. Edge onset identifies an element; integrated intensity supports quantification; and energy-loss near-edge structure (ELNES) can report oxidation, coordination, and bonding when energy calibration, thickness, orientation, and reference spectra are controlled.

The specimen must be thin enough for interpretable transmission, not merely thin enough to form an image. A transmitted electron may scatter inelastically more than once. Plural scattering convolves core edges with the low-loss distribution, redistributes intensity, and can distort fine structure and background. If (I_0) is the integrated zero-loss intensity and (I_t) is the integrated total spectrum, a widely used relative-thickness estimate is

where (t) is specimen thickness and λ is the inelastic mean free path for the material and beam conditions. The ratio (t/λ) is often more defensible than an absolute thickness because converting to nanometers requires a suitable mean-free-path model. Thickness varies across a FIB lamella, so the low-loss spectrum should be paired spatially and temporally with the core-loss data rather than measured once at a convenient location.

Plural events approximately follow Poisson statistics when inelastic events are treated as independent:

This explains why plural-scattering probability grows rapidly with relative thickness. Fourier-log or related deconvolution can estimate a single-scattering distribution when the ZLP and low-loss response are well measured, but deconvolution cannot restore signal-to-noise that was never acquired. It can also amplify artifacts if spectra drift, saturate, truncate the low-loss tail, or use mismatched energy dispersion.

Core-loss analysis depends on background, cross-section, and collection geometry. Before an ionization edge, the decaying background is often modeled over a chosen pre-edge interval, commonly with a power-law form (AE^{-r}). The background is extrapolated under the edge and subtracted; signal is then integrated over a stated window. For a sufficiently thin region and compatible cross-section model, elemental areal density can be estimated as

where (I_k) is extracted edge intensity, α is probe convergence semi-angle, β is collection semi-angle, Δ is the integration window, and σₖ is the partial ionization cross-section. Edge overlap, channel gain, detector point-spread, energy drift, thickness, diffraction, and plural scattering all influence the result. A concentration map without these acquisition parameters is not a portable quantitative measurement.

Spatial resolution is set by more than the STEM probe diameter. Core-loss events with large energy transfer can be highly localized, but inelastic scattering has an energy-dependent delocalization and angular distribution. Low-loss excitations may extend well beyond the nominal probe, while some high-energy edges can support atomic-column contrast in a stable, thin crystal. Channeling, probe tails, scan drift, specimen thickness, detector collection, and the signal extraction model all affect apparent interface width. “Atomic-resolution EELS” describes an achieved experiment under specific conditions; it is not a universal resolution specification for every edge, specimen, or dose budget.

The characteristic scattering angle scales approximately as

in the high-energy small-angle limit. Collection angle therefore changes signal efficiency and the measured momentum-transfer distribution. Too narrow an aperture can reject useful edge intensity and make alignment critical; a wider aperture admits more signal but may increase background or integrate orientation-dependent features differently. Convergence and collection angles, beam energy, energy dispersion, aperture, camera length, and entrance geometry belong with the spectrum because cross-sections and fine structure depend on them.

question[Define element, bonding, dielectric, or thickness question] --> prepare[Prepare representative electron-transparent region]
prepare --> setup[Choose beam energy, dose, dispersion, and angles]
setup --> acquire[Acquire aligned zero-loss, low-loss, and core-loss data]
acquire --> qa{No saturation, drift, contamination, or damage?}
qa -- no --> adjust[Reduce dose or revise preparation and acquisition]
adjust --> acquire
qa -- yes --> thickness[Map t over lambda and assess plural scattering]
thickness --> process[Calibrate energy, model background, deconvolve if justified]
process --> extract[Fit edges or ELNES with references and cross-sections]
extract --> stress{Stable across windows, thickness, and dose?}
stress -- no --> process
stress -- yes --> correlate[Correlate with STEM contrast, EDS, diffraction, and process geometry]
correlate --> report[Report uncertainty, preparation history, and acquisition metadata]

Fine structure is a fingerprint only when references and physics are matched. ELNES reflects transitions from a core level into unoccupied states, so edge onset, peak splitting, and white-line ratios can respond to valence, coordination, crystal field, and bonding. The same features can also change with crystallographic orientation, momentum transfer, thickness, plural scattering, energy resolution, and irradiation. Reference spectra should be acquired or simulated for plausible compounds under comparable conditions, aligned by a stated rule, and tested as alternatives. Assigning an oxidation state from one peak ratio without uncertainty or dose controls is weaker than a model that explains the complete edge shape and agrees with diffraction or chemistry.

Low-loss EELS can probe plasmon energy, interband transitions, and a dielectric response through Kramers–Kronig analysis, but a band-gap number is not obtained by simply drawing a line at the first intensity above zero. The ZLP tail, energy resolution, thickness, surface excitations, retardation effects such as Čerenkov radiation, and guided modes can obscure the onset. Monochromation improves energy resolution while often reducing current or changing dose efficiency. A claimed nanoscale gap or dielectric function should document ZLP removal, collection geometry, thickness, normalization, and the physical model used to separate bulk and surface contributions.

The electron beam and specimen preparation can rewrite the chemistry being measured. FIB milling may implant ions, amorphize surfaces, redeposit material, preferentially thin one phase, or oxidize the cross section during transfer. Protective caps and low-energy final polishing reduce some artifacts but do not guarantee pristine chemistry. During EELS acquisition, radiolysis, knock-on displacement, heating, contamination deposition, reduction, and crystallization may occur. Dose fractionation, fast repeated scans, non-rigid registration, cryogenic methods, lower voltage, and before-versus-after spectra are choices to manage damage; the correct choice depends on the material, edge cross-section, and required spatial resolution.

A spectrum image must be audited as a time sequence as well as a spatial map. Each pixel is acquired at a different time, so energy drift, stage drift, scan distortion, beam-current change, and evolving contamination can masquerade as a compositional gradient. Simultaneous or rapidly interleaved low-loss and core-loss acquisition helps energy alignment and thickness correction. Summing only pixels selected after viewing a noisy map can bias weak-edge claims. Robust analysis declares the region-selection rule, propagates counting uncertainty, compares alternate backgrounds, and tests whether the feature persists in independent scans or orthogonal scan directions.

EELS is most persuasive when it joins complementary signals rather than carrying the interpretation alone. HAADF-STEM supplies mass-thickness and diffraction-sensitive structure; EDS supplies characteristic X-rays with different edge overlaps and sensitivity; diffraction constrains phase and orientation; XPS or XANES provides ensemble chemical-state context; and device geometry constrains which diffusion or reaction pathways are plausible. Together they can distinguish a real interfacial compound from a thickness step, preparation artifact, or beam-induced state.

For semiconductor metrology, the central question is not “can an edge be plotted at atomic sampling?” It is “which composition or electronic-state conclusion survives thickness, plural scattering, background, collection geometry, delocalization, preparation, drift, and dose tests?” Reading EELS through that energy-loss-physics-and-specimen-integrity lens turns a beautiful spectrum image into defensible nanoscale evidence.