Critical-dimension small-angle X-ray scattering measures the average three-dimensional shape of periodic semiconductor structures by transmitting short-wavelength X-rays through a patterned target and recording its diffraction orders. Pitch fixes where those orders appear; the distribution of intensity among them carries the cross-sectional form factor. By rotating the wafer and fitting many orders together, CD-SAXS can constrain line width, height, sidewall angle, multilayer offsets, contact-hole diameter, and other profile parameters without cleaving the wafer. The output is not an X-ray image of one feature: it is the statistically supported periodic ensemble profile whose calculated scattering best explains the calibrated data.

Periodic order positions provide a direct and robust pitch constraint. For a one-dimensional grating with pitch $p$, reciprocal-lattice orders occur at

where integer $m$ labels the order. This relation makes mean pitch and coherent pitch subdivisions among the best-conditioned CD-SAXS outputs when detector geometry and wavelength are calibrated. Missing, split, or satellite orders can reveal superlattices, pitch walking, multiple-patterning asymmetry, or finite correlation. Yet the illuminated target must contain enough coherent repeats, and stage azimuth must align the periodic axis correctly; otherwise order intensity spreads, shifts, or leaves the detector acceptance.

The order intensities encode an electron-density form factor rather than a geometric silhouette. In the kinematic approximation, the complex scattering amplitude is the Fourier transform of electron-density contrast,

For a line whose width changes with height, the amplitude samples the entire function $w(z)$, not only top and bottom CD. Rotating the wafer changes the reciprocal-space trajectory through that form factor, allowing height, sidewall shape, offsets, and buried layers to influence different observations. Material composition and density set contrast; a geometric boundary with little electron-density contrast can remain weak even when it matters electrically.

A profile model converts a finite intensity data set into dimensional parameters. Common models slice the cross-section into height segments, connect control-point widths, or describe trapezoids, multilayer shells, asymmetric line sets, holes, pillars, fins, nanosheets, or high-aspect-ratio channels. Model complexity should grow only when residuals and sensitivity justify it. A simple trapezoid can bias top and bottom CD when the true wall bows; an overly flexible spline can fit noise and make adjacent widths anticorrelate. The best model is the least complex electron-density profile that explains all rotations and orders without systematic residuals and remains stable under plausible background and resolution changes.

Intensity-only detection loses phase and makes the inverse problem non-unique. Different cross-sections can have similar $|F|^2$ over a limited reciprocal-space range, especially when model parameters are correlated. Symmetry assumptions can conceal left–right asymmetry; top CD, wall angle, height, and density can trade against one another. Variable-angle measurements recover additional slices of the three-dimensional transform but do not magically restore all phase information. Competing physical models, multiple starting points, posterior or profile-likelihood exploration, held-out rotations, and sensitivity to parameter bounds are needed before small numerical uncertainties become credible dimensional uncertainties.

The measurement is an ensemble average over a periodic test structure. Thousands or millions of features may contribute coherently or incoherently within the beam footprint. The fitted profile is therefore a density-weighted average, not the shape of a worst-case line, isolated defect, or particular device. Target nonuniformity, finite array size, pattern-density transitions, wafer curvature, edge placement, and beam position affect the result. This averaging can deliver excellent precision for process means while hiding rare excursions. CD-SEM, AFM, or TEM supplies local distributions and defect context; CD-SAXS supplies a nondestructive 3D ensemble constraint. Their measurands should be reconciled rather than expected to match one feature exactly.

Roughness signatures combine several kinds of shape variation. Under restricted assumptions, random displacement or edge variation can attenuate higher orders with a Debye–Waller-like factor such as

but the fitted $\sigma$ need not be pure line-edge roughness. Line-width, height, sidewall-angle, placement, and correlated periodic fluctuations can all redistribute or damp intensity. Instrument resolution and finite coherence also suppress high-order contrast. A roughness metric must state the stochastic model and correlation assumptions; comparing it directly to top-down CD-SEM LER is valid only when both methods respond to the same fluctuation components.

Instrument calibration and data reduction set the dimensional scale. Beam energy, sample-to-detector distance, beam center, pixel geometry, detector distortion, stage rotation axis, sample tilt, exposure normalization, absorption, polarization, background, beamstop masks, and resolution determine where and how strongly orders appear. Direct-beam and reference-standard measurements constrain the reciprocal-space scale. Symmetry between positive and negative orders is a powerful diagnostic for centering and detector response. Dynamic range matters because weak high orders often carry the sharpest profile information; saturated low orders and unqualified stitched exposures can distort the likelihood just as seriously as missing high orders.

st=>start: Define structure, process decision, target, and required uncertainty
design=>operation: Select energy, beam size, rotations, q range, exposure, and test pattern
cal=>operation: Calibrate beam center, distance, wavelength, detector, stage axes, and tilt
acq=>operation: Acquire direct beam, blank, standards, symmetric orders, and replicates
reduce=>operation: Normalize, mask, subtract background, map q, and propagate count uncertainty
model=>operation: Build periodic electron-density profile with absorption and resolution
fit=>operation: Fit all orders and rotations jointly; test complexity and parameter covariance
test=>condition: Stable across starts, rotations, bounds, and alternate models?
revise=>operation: Expand angular/q support or constrain with SEM, AFM, TEM, XRR, or composition
report=>end: Report ensemble profile, model class, uncertainty, and detection limits
st->design->cal->acq->reduce->model->fit->test
test(yes)->report
test(no)->revise->design

CD-SAXS complements rather than replaces neighboring dimensional metrologies. Optical critical-dimension scatterometry is fast and production-proven but depends on optical constants and wavelengths larger than advanced features. CD-SEM localizes top-down edges with high throughput but provides limited buried or full-height information and can be sensitive to charging and edge algorithms. Cross-sectional TEM shows local internal structure destructively over a tiny field. AFM traces accessible surfaces with tip-convolution limits. GISAXS emphasizes surface and thin-film morphology in a reflecting wavefield. CD-SAXS uses transmission and periodic scattering to constrain buried, high-aspect-ratio, and three-dimensional ensemble profiles, making it especially valuable for FinFET, gate-all-around, 3D NAND, DRAM, TSV, contact-hole, and multiple-patterning test structures.

A production CD-SAXS report records target layout and finite size, pitch family, materials and assumed electron densities, beam energy and size, detector and stage calibration, wafer rotations and azimuth, exposure and normalization, masks, absorption and background, resolution model, profile parameterization, parameter bounds, objective function, count statistics, covariance, alternate-model tests, replicates, and orthogonal validation. It distinguishes precision within the chosen model from uncertainty that includes model discrepancy. With those controls, critical-dimension small-angle X-ray scattering becomes a periodic-ensemble-form-factor-and-model-identifiability lens.