An ellipsometer measures a polarisation change, not a thickness. What separates a defensible result from a convenient one is the layer stack you assumed, the parameters you allowed to correlate and the independent check you did or did not run.
A student reports a film thickness to four significant figures because that is what the fitting software displayed. Nobody asks which model produced it, what was fixed, what floated, or whether a different starting guess would have produced a different number. The instrument is excellent and the measurement is precise. Whether it is correct depends entirely on a set of choices that were made in a dialogue box and never recorded.
An ellipsometer sends polarised light onto the sample at an oblique angle and measures how the polarisation state changes on reflection. The two quantities it reports, conventionally called psi and delta, describe the amplitude ratio and the phase difference between the two polarisation components. Because they form a ratio, the measurement is self referencing: it does not depend on the absolute intensity of the source, which is why ellipsometry is so precise and so stable over time.
It is also why it cannot give a thickness directly. Psi and delta at each wavelength are the consequence of a whole optical stack: substrate, interfacial layers, film, roughness, ambient. To recover a thickness or a set of optical constants you build a model of that stack, calculate what psi and delta it would produce, and adjust the free parameters until the calculation matches the measurement. The output is not a measurement of the film. It is the film that best explains the measurement, given the model you chose. Every failure mode below is a failure of that model.
The substrate is the foundation and it is where the first errors enter. Measure the bare substrate before anything is deposited on it, fit it, and then hold it fixed. Silicon carries a native oxide of a few nanometres that is not optional in the model; leaving it out shifts everything above it. Glass and quartz vary between suppliers and between batches, so a library entry for fused silica is not a safe substitute for measuring the actual substrate you are using.
Above the substrate, the film layer needs an optical dispersion model. The choice depends on what the film does with light.
Point by point inversion, where the optical constants are extracted independently at every wavelength, should come after a parametric fit has converged and only to refine it. Run it first and you fit the noise, producing optical constants that wander and a thickness that is meaningless.
For a single transparent film measured at one angle, thickness and refractive index are not independent. A slightly thicker film with a slightly lower index produces almost the same optical path and almost the same data. The fitting routine will happily converge, report a small mean squared error, and hand you a pair of numbers that could equally well have been a different pair.
The software knows this and reports it in the correlation matrix, which is the part of the output nobody looks at. When two fitted parameters are correlated close to unity you are effectively fitting one parameter and reporting two. The fixes are all about adding information that breaks the degeneracy.
Surface roughness is conventionally handled as a thin effective medium layer of fifty per cent film and fifty per cent void. That is a modelling convenience, and it works when the roughness is small compared with the wavelength. It stops working when it is not, and the way it fails is instructive: the fit still converges, the roughness layer grows to an implausible value, and it starts absorbing other errors in the model.
So put a number on the roughness before you fit it. An atomic force microscope scan of the same sample takes minutes and tells you whether a two nanometre roughness layer is physical or a fudge. If the surface is genuinely rough on the scale of the wavelength, light is scattered out of the specular beam and depolarised, and the instrument will report depolarisation. That is a signal to stop trusting a simple model rather than to add another layer.
Grading is the related problem. A film whose composition or density changes through its thickness is not one layer, and forcing it into one produces optical constants that belong to no real material. Graded models exist and should be used where the deposition process is known to produce them, for example a reactive sputter run where the oxygen flow was ramped. Two thin sublayers with a shared dispersion and different void fractions is often enough to reveal whether grading is present.
Two mechanical issues cause more trouble than either of these on a first measurement. Light reflected from the back surface of a transparent substrate returns into the detector and corrupts the data; roughen the back, tape it, or index match it. And sample tilt changes the effective angle of incidence, so use the alignment routine on every sample rather than assuming the stage has not moved.
A low mean squared error is necessary and nowhere near sufficient. Enough free parameters will fit anything. These are the checks that distinguish a result worth defending.
A thickness quoted without its model is not a measurement, it is a claim. The minimum record is the angle set, the spectral range, the substrate file and where it came from, the dispersion model for each layer, which parameters were fitted and which were fixed and to what values, whether a roughness layer was included and how thick it came out, the mean squared error, and any independent check. That is a short paragraph in a logbook and it is the difference between a result that survives review and one that has to be repeated.
For a shared instrument, put the standard models in a locked template so that routine users cannot silently change the substrate or add layers. Give experienced users their own template. Most bad ellipsometry in a multi user facility comes from a model that drifted from one project to the next without anyone noticing.