Your second sentence is the take home message. From my experience, I would not depend only on ellipsometry to determine thickness as an a priori unknown value. Reliable determination of this value for a film depends on knowing refractive index and surface roughness values accurately for both the film and its substrate. I might instead attempt to cross calibrate ellipsometry measurements with a less ambiguous (more accurate and precise) method to measure thickness directly. Sputter profiling with XPS comes to mind immediately. Presuming that, at each film thickness, all other parameters stay the same (refractive index and surface roughness), you would have a starting point to use ellipsometry to obtain film thickness. The best situation is to model the film holding all other parameters constant at any film thickness. Your only variable is film thickness. It will be a false value from ellipsometry. Plot this false value versus the true value determined independently to obtain a calibration curve.

If you have access to an AFM, I would also use that as a second way to measure thickness. 

Thank you Jeffrey and Blanca for your suggestions.

In our lab we do not have AFM, we measure roughness with WLI and that is what I used for that purpose.

Jeffrey, your idea of fixing all parameters and creating a calibration curve is what I want to do but I am not there yet because I do not know what the right parameters are for this material.

I have film thickness results from samples using TOF-SIMS and roughness from WLI as mentioned before. So when I am developing my model I have fixed a layer of roughness and film thickness yet not able to get optical properties consistently. For instance, I am using a Drude model plus a Lorentz harmonic oscillator to model the thinnest sample. Plasma frequency for Drude is taken from literature so I am left with the HO only, which I model and fine tune those parameters to get the lowest GOF. Then I move to my next sample, which is about the same thickness and my model does not work anymore. 

I found also that roughness measurements that I am getting from WLI have a determinant impact in my model, meaning that if I measure roughness in different locations of the same sample and those measurements are different then my model parameters change significantly and my model is not consistent anymore.

Currently I am creating a user defined material for my model and I decide to use the combination of Drude plus Lorentz HO, but I have to my disposal many more options that potentially would give me similar results yet parameters are different and physical meaning not to clear to me since they are buried in mathematical calculations. Yet I am open for ideas that would ease this process.

Thank you for your time and any advise would be greatly appreciated. 

First a disclaimer. I am by no means conversant in ellipsometry. I know enough to shoot myself in the foot as the saying would go. I already see that you are far ahead in the theoretical background. As an experimentalist, I think of a simpler approach. This is not one that involves spectroscopy as you may be doing but one that only takes a point-by-point value of intensity at a given angle and wavelength. It has only two parameters for each of two layers (film + substrate), index of refraction n and roughness r. The one remaining is the film thickness t. This is five parameters. Fix n and r for the substrate. That leaves n, r, and t for the film. You are measuring r and t. Rather than trying to model the film through theory to obtain n (likely as a function of wavelength), try making a correlation plot. Pick one value of wavelength and angle. Fit the intensity with a value of n for a given value of r and t. What type of correlations do you obtain from this? I am thinking here of a "seat of the pants" scatter plot. This quick and dirty analysis could perhaps help you see whether your data set is too small or whether something else is undefined that disallows you from doing the more extensive correlations that you are trying to do. IOW, when your spectroscopy is to fit over the entire wavelength and angle range, then it must certainly fit at a fixed wavelength and angle to start.

Presumably, you are measuring ellipsometry over the same region where you are measuring WLI and TOF-SIMS? Also, the WLI makes its own assumptions about n for the surface film to obtain a value of r. Presumably, this is not a case where the WLI values of r are confused by using the wrong value of n for the film on a location by location basis. Here, the recommendation to use a contact metrological tool such as profilometry or AFM comes to mind.

Otherwise, you are finding what I have learned second hand and otherwise guessed. Ellipsometry is great for characterizing one parameter on an otherwise homogeneously well-defined film. It falls apart or becomes unusually ambiguous for films that are complex.

may be your sample have large pin holes, try first to measure roughness by AFM, and the better way to measuring optical constants, deposit your material on silicon substrate.

One approach is to grow a very thick layer on glass, and use ellipsometry to determine n and k for that material. You should be consulting with your equipment manufacturer or others to do that correctly.

Assuming that material deposits the same every time (which is an assumption), then you should be able to set up models on either Si wafers, or on say Al coated wafers. In either case the layer coated upon must be opaque, or another way to say, optically thick enough not to see through. 

Using such a test wafer one should be able to easily measure thickness of films using the n and k previously determined.  There are some limitations, for example, at very thin layers your material may be undergoing island growth and the model will not be valid.

You could back up  and confirm your results with other techniques as others have suggested. I like FIB-SEM cross-sections for visual confirmation.  But that technique is not always convenient or readily available.

Hello,

when you create the model, you need to insert a first layer with the substrate data (eg the glass or silicon n e k) and then insert the layer of SnO2, where you set the parameters to fit, and as last layer you can use a roughness (if you have a rough sample).

I guess yoಉ need to have other data of the sample like refractive index to correctly estimate thickness with ellipsometry data.