**Every mapped parameter comes from a declared optical stack model.** The forward model uses Fresnel coefficients and propagation through each layer to predict the polarization response. For layer $j$, a phase thickness contains $$ \beta_j=\frac{2\pi}{\lambda}\tilde n_j d_j\cos\theta_j $$ where $d_j$ is physical thickness, $\tilde n_j$ is complex refractive index, and $\theta_j$ is the complex refraction angle implied by Snell’s law. Multiple reflections make the spectrum sensitive to phase and absorption. Interfaces, graded layers, anisotropy, and roughness modify the transfer calculation. Model construction should follow known process history and independent evidence. A plausible film may need an interfacial oxide, composition gradient, surface roughness layer, native contamination, or absorbing substrate. Adding every imaginable layer is not safer: weakly constrained layers trade thickness and optical constants, making the inverse problem ill-conditioned. Begin with the simplest physically defensible stack, examine residual structure, and add complexity only when it is identifiable and improves withheld data or orthogonal agreement. Surface roughness is often represented by an effective-medium layer mixing film and void. Its fitted thickness is a model parameter, not automatically the root-mean-square height from atomic-force microscopy. Correlation length, slope, lateral scale, and scattering are largely absent from a simple effective-medium approximation. When roughness is large relative to wavelength or creates significant diffuse scattering and depolarization, specular ellipsometry alone is insufficient. Ultra-thin interface layers and optical constants are strongly correlated. Fixing validated constants can stabilize thickness mapping; freeing every optical term locally can convert noise into composition. A hierarchical fit can estimate shared dispersion from representative spectra, then map only identifiable local parameters. |Mapping strategy|What varies by coordinate|Principal benefit|Main identifiability risk|Required diagnostic| |---|---|---|---|---| |Fixed optical constants, local thickness|One or several layer thicknesses|Stable high-throughput uniformity map|Real composition or density change is forced into thickness|Spectral residuals and representative free-dispersion fits| |Local thickness plus limited dispersion parameter|Thickness and one process-sensitive optical term|Separates some density/composition variation|Strong thickness–index covariance|Parameter correlation and profile likelihood| |Multi-angle local fit|Same stack fit jointly across angles|Adds sensitivity and tests geometry consistency|Angle-dependent footprint samples different regions|Registered footprints and angle calibration| |Imaging ellipsometry|Pixel- or superpixel-level model parameters|High spatial density over a field|Pixel calibration, focus, angle spread, low signal|Flat-field, polarization, and spatial-resolution validation| |Global or hierarchical wafer fit|Shared optical constants with local thicknesses|Uses all sites to stabilize common physics|Shared parameters can hide real spatial optical variation|Held-out sites and comparison with unconstrained regions| **Optical dispersion must be physical over the measured spectral range.** In a transparent region, a Cauchy-type relation may compactly describe refractive index, but it should not be extrapolated through absorption or used as a microscopic band-structure model. Absorbing films require a causal dielectric function or oscillator model suited to the material and energy range. Kramers–Kronig consistency links real and imaginary response; flexible point-by-point functions need regularization and should not generate negative absorption or nonphysical discontinuities. The chosen spectral window controls parameter sensitivity. Below a film’s absorption edge, interference can constrain optical thickness but leave physical thickness and refractive index correlated. Near electronic transitions, spectral shape helps determine dispersion and composition but also introduces resonance, roughness, and broadening parameters. At energies where substrate or ambient absorption dominates, information about buried layers may collapse. Multi-angle measurements alter field penetration and p/s sensitivity, often improving identifiability. However, changing incidence angle elongates and rotates the footprint and can sample different material on a nonuniform wafer. Joint fitting assumes the same local stack at all angles. Registration error must be smaller than the spatial scale of variation, or the added “information” is a mixture of locations. Parameter covariance should be measured, not inferred from a smooth map. Covariance, profile likelihood, bootstrap, or synthetic recovery can expose ambiguity. Optimizer errors are unreliable when the model is wrong, parameters sit on bounds, or calibration uncertainty is omitted. A sensitivity matrix can be written $$ J_{ab}=\frac{\partial y_a}{\partial p_b} $$ for measured observables $y_a$ and parameters $p_b$. Nearly dependent columns of $\mathbf J$ indicate parameters that the dataset cannot separate. Add independent angles, wavelengths, reference data, or physical constraints; do not merely report more decimal places. **Mapping quality is diagnosed spatially through residuals and parameter behavior.** A scalar mean-squared-error value summarizes fit mismatch but hides wavelength structure and compensation among $\Psi$ and $\Delta$. Save residual spectra at every site or at least representative and worst-case locations. Map residual norm, degrees of freedom, convergence status, parameter bounds, uncertainty, and correlation alongside thickness or optical constants. Residual patterns can identify a missing layer, angle offset, backside reflection, depolarization, or calibration error. Rings and stripes may follow process variation, wafer bow, autofocus, stage motion, or detector stitching. Repeat scan direction and mounting before assigning them to process physics. Goodness of fit cannot establish uniqueness. Use physical bounds, causal dispersion, independent measurements, and held-out tests. When models fit comparably, report the ambiguity or retain only quantities stable across them. Neighbor seeding can propagate a wrong local minimum, while smoothing can erase edge or die structure. Use independent restarts, retain unsmoothed estimates, and declare regularization or interpolation. Spatial outliers deserve classification rather than automatic deletion. They may be particles, scratches, mixed footprints, focus failures, backside contamination, true process defects, or model breakdown. A robust rule should use spectral residuals, repeatability, image or reflectance context, and neighboring behavior. Report exclusion masks and counts so uniformity metrics can be reproduced. **Anisotropy and depolarization define the boundary of conventional mapping.** The scalar ratio $r_p/r_s$ assumes no p-to-s polarization conversion and a nondepolarizing sample. Anisotropic crystals, oriented polymers, slanted columns, textured films, magnetic response, patterned structures, or off-axis geometry can require a Jones-matrix or generalized ellipsometry description with cross-polarization coefficients. Depolarization occurs when the detector averages incoherent polarization states from thickness variation, roughness, patterned mixtures, finite angular spread, or multiple backside paths. A Mueller-matrix measurement can quantify depolarizing behavior that a simple $\Psi,\Delta$ model cannot represent. Fitting depolarized data with an isotropic stack often maps the unmodeled physics into false roughness, thickness, or optical constants. Generalized and Mueller-matrix ellipsometry are distinct extensions with richer observables and calibration demands. For routine mapping, a practical boundary test is to measure depolarization or selected off-diagonal terms at representative sites. If they exceed the validated tolerance, switch models or classify the site as outside the conventional method’s domain rather than force a scalar fit. Patterned wafers may violate lateral homogeneity even when the pattern is much smaller than the footprint. If the pitch is far below wavelength and conditions support homogenization, an anisotropic effective-medium model may work. When diffraction orders propagate or critical dimensions influence the response, rigorous coupled-wave analysis or another scatterometry model is needed. A blanket-film ellipsometry recipe cannot be transferred to product patterns solely because their average reflectance looks similar. Transparent substrates introduce backside reflection. A coherent backside beam can produce spectral fringes; an incoherent contribution can depolarize or bias the front-stack response. Roughening or masking the backside, wedged substrates, spatial filtering, coherence modeling, or explicit backside optics may be required. The treatment must remain consistent across sites, especially if substrate thickness or backside condition varies. **Spatial statistics must respect sampling, boundaries, and measurement uncertainty.** Wafer uniformity is often summarized by range, standard deviation, percent nonuniformity, radial profile, or site-to-site difference. Every metric needs a declared site set, edge exclusion, center convention, weighting, and denominator. Range is highly sensitive to a single bad fit; standard deviation mixes true spatial variation with measurement noise; percent metrics become unstable when the mean approaches zero. Separate repeatability from wafer variation using repeated sites, repeated maps, or a nested measurement design. If $s_{obs}^2$ is observed site variance and $s_{meas}^2$ is repeatability variance under compatible assumptions, a process component may be estimated from their difference, but negative or spatially varying results require a fuller model. Drift and correlated footprints violate simple independent-noise subtraction. Radial averaging can hide azimuthal signatures and defects. Polynomial or Zernike summaries compress low-order variation but are not physical process models; spatial correlation methods require a grid that resolves the relevant length scale. Interpolation creates values where no spectrum was measured and cannot exceed footprint resolution. Validate the interpolation and do not bridge notches, bevels, pattern boundaries, or excluded sectors. Control limits should reflect uncertainty and model validity. Keep separate flags for acquisition failure, model failure, parameter excursion, and spatial-rule violation so process control does not react to an optical artifact. ```flowchart Define the film parameter, spatial scale, wafer coordinates, and decision limit -> Choose wavelengths, angles, footprint, grid, references, and exclusions -> Build the simplest process-informed optical stack and dispersion model -> Calibrate polarization, wavelength, angle, stage, focus, and backside handling -> Acquire registered spectra with repeated reference and wafer sites -> Fit parameters with bounds, covariance, restarts, and residual retention -> Map parameters, uncertainty, residuals, convergence, and exclusion masks -> Test alternate models, spatial artifacts, repeatability, and held-out data -> Validate representative sites using thickness or composition references -> Release process metrics only inside the model and sampling validity domain ``` **Traceability and reproducibility require reference artifacts plus model provenance.** A thin-film reference can check tool stability and bias, but its value depends on material, thickness, substrate, aging, cleanliness, and reference method. NIST intercomparisons have shown that instrument and algorithm differences can create systematic thickness differences even on nominally simple oxide stacks. A reference controls the measurement chain only when its uncertainty, environmental condition, and optical model are documented. Daily or lot-level checks should monitor $\Psi$ and $\Delta$ or Mueller elements directly, not only fitted thickness. A stable thickness can hide compensating drift in angle and optical constants. Track wavelength calibration, angle calibration, polarizer and analyzer state, compensator response, detector linearity, source spectrum, focus, stage coordinates, and reference residuals. Control charts should distinguish abrupt maintenance changes from gradual source or contamination drift. For each map, preserve raw spectra, coordinates, timestamps, tool state, recipe, model graph, optical-constant source, parameter bounds, initialization, software version, fit results, covariance, residuals, masks, interpolation, and summary code. Report enough metadata to reproduce both the parameter map and its diagnostics. A static image without its model and site table is not a metrology record. Validation should challenge the map at representative center, edge, high, low, and poor-fit sites. Cross-sectional microscopy, x-ray reflectometry, profilometry, reflectometry, composition analysis, or calibrated step structures can test different parts of the result. These methods have different footprints and model assumptions, so compare forward-predicted observables or carefully matched regions rather than expect exact agreement by default. The durable way to interpret spectroscopic ellipsometry mapping is through a polarization-observable-footprint-stack-model-identifiability-diagnostic-spatial-statistics-and-traceability lens.