Ocd (optical Critical Dimension)
Optical critical-dimension scatterometry infers the average geometry of a periodic semiconductor pattern from how that pattern changes reflected or diffracted light. The tool may report linewidth, height, sidewall angle, corner rounding, film thickness, and overlay-related parameters without cutting the wafer, but those values are not read directly from an image. They are the parameters of an electromagnetic model whose simulated signature best explains the measured spectrum, angle response, polarization state, or diffraction orders.
The optical signature is a collective response of the modeled structure. Depending on the instrument, observables may include reflectance, transmittance, ellipsometric $\Psi$ and $\Delta$, Mueller-matrix elements, or resolved diffraction efficiencies as functions of wavelength, incidence angle, azimuth, and polarization. For a simple grating, propagating orders satisfy a relation of the form
where $p$ is pitch and $m$ is diffraction order. When pitch is subwavelength, higher orders may be evanescent in the far field, yet the zero-order polarization and spectral response still carry profile information through electromagnetic coupling within the grating.
A forward solver turns an assumed profile into predicted data. Rigorous coupled-wave analysis, finite-element, finite-difference time-domain, or integral-equation methods solve Maxwell’s equations for the parameterized stack. The parameter vector may contain top and bottom CD, height, sidewall angle, corner radius, undercut, residual layer, pitch, overlay, film thicknesses, and complex refractive indices. Discretization order, mesh, Fourier harmonics, boundary conditions, material anisotropy, and convergence tolerance must be tight enough that numerical error is small relative to the measurement requirement.
The inverse problem selects parameters by comparing simulation with measurement. A covariance-weighted objective can be written
where $\mathbf{y}$ is the measured signature, $\mathbf{f}(\mathbf{p})$ the forward model, and $\mathbf{\Sigma}$ the measurement covariance. A precomputed library searches a discrete parameter grid; regression iteratively updates parameters; surrogate or machine-learning models approximate the forward or inverse map. All three approaches inherit the same physics and identifiability limits, even when their runtimes differ dramatically.
| OCD element | What it contributes | Primary benefit | Failure mode to control |
|---|---|---|---|
| Spectral reflectometry | Intensity versus wavelength | Fast broadband sensitivity | Limited polarization information and source drift |
| Spectroscopic ellipsometry | Polarization amplitude and phase | Strong film and profile sensitivity | Optical-constant and depolarization model errors |
| Angle-resolved measurement | Signature versus incidence or collection angle | Adds independent geometric sensitivity | Angular calibration, footprint, and stage alignment |
| Mueller-matrix measurement | Full polarization transfer | Detects anisotropy, asymmetry, and depolarization | More calibration terms and larger inverse model |
| Periodic target design | Controlled pitch, stack, and orientation | High signal and repeatable process monitor | Target-to-device bias and nonrepresentative loading |
| Cross-metrology reference | CD-AFM, CD-SEM, TEM, or X-ray constraints | Tests absolute accuracy and model form | Different averaging volumes and measurand definitions |
Identifiability matters more than the number of fitted digits. The local sensitivity matrix
shows how each optical datum responds to each parameter. Nearly collinear columns mean two profile changes produce similar signatures; linewidth and height, film thickness and optical constants, or sidewall angle and corner rounding may become strongly correlated. Under a locally linear, correct-model approximation, parameter covariance is often estimated as
A singular or ill-conditioned matrix signals that the recipe does not independently constrain all requested parameters. These parameter correlations must be reported rather than hidden by fixing one correlated input to an incorrect nominal value, which can make the remaining outputs repeatable and biased.
Residuals test model adequacy rather than merely fit quality. Random residuals consistent with measurement noise support the chosen model locally. Wavelength-correlated, polarization-specific, or angle-dependent residuals point to missing layers, incorrect optical constants, target asymmetry, roughness, depolarization, numerical error, or calibration drift. A small scalar mean-square error can conceal structured residuals across thousands of points. Recipe acceptance should therefore include residual plots, alternate parameterizations, convergence from multiple starting points, and holdout conditions not used in fitting.
A useful OCD target is designed in sensitivity space before it is printed. For optical channel $i$ and profile parameter $p_j$, the Jacobian element $J_{ij}=\partial f_i/\partial p_j$ is the local fingerprint plotted above. A large derivative gives sensitivity; a distinct spectral shape gives separability. Two strong but nearly proportional columns are still a bad recipe because the fit can exchange those parameters with little change in residual.
Singular-value decomposition makes that failure quantitative:
A small $\sigma_{\min}$ or large condition number identifies a weak combination of profile parameters. The remedy is experimental, not algorithmic: change pitch or stack, add an azimuth, angle, polarization or wavelength band, or constrain one parameter with cross-metrology. Regularization may stabilize a number, but it cannot manufacture information that the optical channels never measured.
That target-and-recipe co-design is the economic moat. A qualified library, converged solver, reference hierarchy and drift monitor convert a fast optical measurement into actionable geometry; without them, the same instrument produces precise model-dependent numbers whose absolute bias remains unknown.
The first loop is the one that decides whether the measurement is possible at all. If two parameters are not independently observable in the chosen optical channels, no amount of fitting separates them later — the target or the optics have to change instead.
The reported profile is an optical effective average. The illuminated spot covers many nominally periodic features, so extracted dimensions represent the model-equivalent response of that ensemble. Line-edge roughness, line-width roughness, pitch walk, stochastic defects, local loading, and across-spot gradients can broaden or depolarize the signature without mapping one-to-one onto a trapezoid parameter. OCD provides excellent high-throughput process averages; it does not replace local imaging when the question concerns an individual bridge, break, stochastic contact failure, or extreme tail of a distribution.
Target and device equivalence must be demonstrated. Large periodic gratings provide strong optical sensitivity but can print, etch, clean, or polish differently from product structures because of pitch, density, neighborhood, stack, or pattern orientation. Correlation to electrical or cross-sectional device measurements establishes a target-to-device offset only over the validated process space. A stable correlation can fail after a material, resist, etch chemistry, optical constant, or design-rule change. Product-like targets and periodic recertification reduce that transfer risk.
Optical constants are coupled model inputs, not universal handbook numbers. Refractive index and extinction coefficient depend on wavelength, composition, density, crystallinity, temperature, and sometimes thickness or anisotropy. Fitting geometry and optical constants simultaneously can create severe covariance. Independent film-stack ellipsometry, witness wafers, constrained dispersion models, and physically reasonable bounds help, but the reference films must represent the patterned process. Native oxide, residue, hard mask, sidewall polymer, and buried interfaces can matter even when individually thin.
Precision, sensitivity, and accuracy answer different questions. Repeat measurements may show subnanometer precision because the optical signal is stable, while absolute accuracy remains limited by systematic calibration, model discrepancy, parameter correlations, optical constants, target nonuniformity, and reference uncertainty. NIST uncertainty work emphasizes propagating both measurement noise and systematic effects and visualizing correlated profile uncertainty. A production control limit can legitimately use a precise relative metric, but it should not be presented as traceable absolute geometry without suitable references and an uncertainty budget.
The strongest OCD recipe is not the one that returns the most profile parameters; it is the one whose target, optical channels, forward model, residuals, correlations, and reference measurements make the needed parameters identifiable and traceable. That is the forward-model-identifiability-and-traceability lens.