Ellipsometry is one of those techniques almost nobody outside a materials lab has heard of, and almost every advanced manufacturing line quietly depends on. It is how you establish how thick a layer is when that layer is a few tens of nanometres thick — too thin to see, too thin to touch, and far too important to guess at. On 26 August 2026 the University of Toledo published a profile of a physics doctoral student who has spent most of his degree attacking the slowest part of it.

The student is Alex Bordovalos, and the slow part is not the measurement itself. Modern instruments collect a spectrum in seconds. What takes the time is turning that spectrum into a number a process engineer can act on, and that step still leans on a trained physicist sitting at a workstation adjusting a mathematical model by hand until it agrees with the data.

“It’s not like you can take a ruler and see it,” Bordovalos told UToledo News. “You need specialized tools.” His supervisor put the industrial consequence plainly. “Alex is doing great work that addresses a major bottleneck in industry and research,” said Dr. Nik Podraza, professor of physics and interim dean of the College of Natural Sciences and Mathematics at Toledo, who advises him.

That word — bottleneck — is the whole story, and it is worth being precise about where the bottleneck actually sits. It is not in the optics, the detector, the sample stage or the wavelength range. It sits in analysis, and analysis is exactly the sort of pattern-heavy, judgement-laden work that machine learning has been eating into across the physical sciences for the past five years.

This article walks through what ellipsometry measures, why its interpretation step became a production constraint, what Bordovalos and his co-authors actually published in the Journal of Applied Physics in September 2025, the numbers behind the training set, what still has to be proved on real cadmium telluride solar cells, and what any of it means for a business that buys measurement software rather than writes it.

What UToledo Said About the Ellipsometry Work

ellipsometry ai tools industry bottleneck b microscope arm and tube

The university’s announcement is short, and it is worth separating what was stated from what has been inferred elsewhere in this piece.

The claim being made

UToledo says Bordovalos is developing artificial intelligence technology to automate the most time-consuming and labour-intensive element of spectroscopic ellipsometry — interpreting raw data into measurements that mean something. That is the claim. It is a claim about analysis throughput, not about a new optical method.

The supervisor’s framing

Podraza, who is both Bordovalos’s advisor and interim dean of the college, described the ellipsometry work as addressing “a major bottleneck in industry and research” and said the tool “promises huge decreases in analysis time in measurements of materials that may be many times thinner than a human hair.” He named the beneficiaries as manufacturers in the semiconductor and energy sectors.

The origin story

The idea came out of an industry internship built into Toledo’s Professional Science Master’s in Photovoltaics, which Bordovalos completed in 2022. Working inside a solar cell manufacturer, he saw how hard it is to spot non-uniformities on a production line when the features in question are measured in nanometres.

What has not been claimed

No commercial product has been announced. No accuracy or speed figure has been attached to the doctoral tool in the university’s release. No licensing partner, spin-out or funding round is mentioned. The published work behind the profile is a validation study on amorphous silicon films, which is a narrower thing than a production-ready analyser, and this article treats it as such.

What Spectroscopic Ellipsometry Actually Measures

ellipsometry ai tools industry bottleneck c crank wheel with handle

Before the automation argument makes sense, ellipsometry itself has to be clear, because the reason analysis is hard is baked into how the measurement works.

Polarisation, not brightness

Ellipsometry shines light at a sample at an angle and measures how the polarisation state of that light changes on reflection. It does not measure brightness the way a camera does. Two quantities come back for every wavelength, conventionally written as psi and delta, describing the amplitude ratio and the phase shift between the two polarisation components.

Ellipsometry measures the sample indirectly

Here is the crux. Those two quantities are not thickness, and they are not refractive index. They are a signature produced by whatever stack of layers the light passed through and bounced off. To get thickness or optical constants out, you must propose a model of the sample — how many layers, in what order, made of what — and then adjust that model until its predicted spectrum matches the measured one.

Why the ellipsometry inversion is hard

The relationship runs one way easily and the other way badly. Given a structure, predicting the spectrum is straightforward physics. Given a spectrum, recovering the structure is an inverse problem with many plausible answers, and ellipsometry inherits every difficulty that comes with that: local minima, parameter correlation, and results that depend on where the fit was started.

Where the human sits in the loop

That is why an experienced analyst is currently load-bearing. Somebody has to decide whether a surface roughness layer belongs in the model, whether an interfacial oxide needs its own entry, which dispersion formula suits the material, and what starting values will not send the regression into a nonsense solution. Ellipsometry is powerful precisely because the model carries so much information — and expensive for exactly the same reason.

Why Ellipsometry Data Analysis Became the Bottleneck

ellipsometry ai tools industry bottleneck d bin full of small cubes

The bottleneck is not a recent development. It is a structural consequence of instruments getting faster while interpretation stayed manual.

Ellipsometry acquisition outran interpretation

A modern mapping ellipsometer will collect spectra at dozens or hundreds of points across a wafer or a plate of glass in the time it takes to make a cup of tea. Every one of those points is a separate inverse problem. The published Toledo study measured an 81-point map from each of its samples, which is a modest map by industrial standards and already 81 fits.

The expertise is scarce and slow to build

Fitting analysis has historically been done manually, guided by intuition, and it is extremely time consuming because a realistic model needs many variables to express both structure and optical response. That intuition takes years to develop. A production line cannot recruit its way out of that, and a research group cannot scale it.

The literature has been saying so for years

This is a well-documented complaint rather than a Toledo-specific one. Work on fully automated analysis notes that the fitting step is the main reason inexperienced users are reluctant to adopt ellipsometry at all, and that robust software which minimises the modelling burden is the main obstacle to wider adoption.

The commercial cost of the delay

The shape of the problem, in one line
The instrument is fast, the interpretation is slow, and the interpretation is the part that decides whether a coating run gets released or scrapped. Every hour spent fitting is an hour of product sitting in quarantine.

On a manufacturing line, a measurement that cannot be interpreted quickly is not really a measurement — it is a record you consult after the fact. That is the difference between metrology used for process control and metrology used for post-mortems, and it is the gap the Toledo work is aimed at.

Inside the Published Ellipsometry Neural Network Study

ellipsometry ai tools industry bottleneck e two arrows short and tall

The doctoral profile is journalism. The underlying evidence is a peer-reviewed paper, and it is specific enough to describe properly.

The peer-reviewed ellipsometry paper

Bordovalos is first author on “Artificial neural network assisted spectroscopic ellipsometry data analysis of hydrogenated amorphous silicon thin films,” published in the Journal of Applied Physics on 3 September 2025 (volume 138, issue 9, article 094902). His co-authors are Venkanna Kanneboina, Prabin Dulal, Balaji Ramanujam, Ambalanath Shan and Nikolas J. Podraza. The work acknowledges support from the Air Force Research Laboratory.

Two networks, two different jobs

The architecture is the interesting part, and it maps neatly onto the two things a human analyst does. One series of networks is trained to generate the structural model — how many layers, and which ones. A second series is trained to generate the optical and structural parameter values that populate that model. The tool is not replacing the physics; it is replacing the guesswork about where to start.

The samples

Three films were used for validation. Two were hydrogenated amorphous silicon deposited on native-oxide-coated crystalline silicon wafers. The third was the same material deposited on soda-lime glass. An 81-point map of spectra was measured from each sample and analysed twice: once with traditional least-squares regression, once with the network-assisted route.

The result that matters

Traditional regression determined that the film on soda-lime glass required an optically distinct surface layer in its structural model, which the films on crystalline silicon did not need. The network-assisted method reached the same conclusion independently — it identified the substrate and detected the presence of that surface layer from the experimental data. Agreement on a structural judgement is a stronger result than agreement on a number.

What the ellipsometry outputs are used for

The parameters the networks produce are not treated as final answers. They are fed into the least-squares regression as starting values, which then produces the same quality of fit as the traditional analysis, generally in less computational time and — the load-bearing phrase — without human-defined initial parameters.

SampleFilmSubstrateMap pointsSurface layer needed
1a-Si:HNative-oxide-coated c-Si wafer81No
2a-Si:HNative-oxide-coated c-Si wafer81No
3a-Si:HSoda-lime glass81Yes

The Numbers Behind the Ellipsometry Training Set

ellipsometry ai tools industry bottleneck f upright disc in holder

The doctoral tool described in the UToledo profile goes further than the published study, and the scale of its training set is the detail Bordovalos chose to volunteer.

Half a million simulated ellipsometry spectra

“I’m working with half a million simulations,” he said. “The simulations cover all these different possibilities. For example, I’ll start with the thickness that I expect, and I might vary it by about 50% in either direction.” That is the entire method in two sentences: build a synthetic universe of plausible samples, compute what each one would look like, and learn the mapping backwards.

Why simulation rather than measured samples

You cannot label real data at this scale, because labelling it would require doing the very analysis you are trying to automate. Simulation sidesteps that completely. The physics that predicts a spectrum from a structure is well understood and cheap to run, so the ground truth is exact by construction. That is why nearly all of the training data here is synthetic.

The sweep window

Varying an expected thickness by roughly half in either direction gives a search window running from 50% to 150% of nominal. It sounds generous until you remember that a real coating run can drift, a new recipe can miss, and a network that has only seen the happy path is worthless precisely when you need it.

Thickness sweep window used when generating the simulations (nominal = 100%)
Lower bound of the sweep 50%
Expected nominal thickness 100%
Upper bound of the sweep 150%

The ratio nobody states out loud

Set the half million simulations against the published validation set and the asymmetry is stark. Three samples at 81 points each is 243 measured spectra. Half a million simulated spectra against 243 measured ones is roughly 2,058 synthetic examples for every real one. That ratio is normal for this style of work, and it is also the single biggest reason to be careful about the final claim.

Measured spectra in the published validation study (243 in total)
Sample 1, film on a silicon wafer 81
Sample 2, film on a silicon wafer 81
Sample 3, film on soda-lime glass 81
All three samples combined 243

Traditional Regression Versus ANN-Assisted Ellipsometry

It helps to lay the two workflows side by side, because the difference is narrower and more useful than “AI replaces the physicist” would suggest.

The ellipsometry regression stays

Nothing in the published approach discards least-squares regression. The final fit is still a regression fit, and it is still judged by the same fit quality. What changes is who supplies the model and the starting point.

The expertise moves upstream

In the traditional route, a person encodes their experience into the ellipsometry analysis every single time, sample by sample. In the assisted route, that experience is encoded once, into the design of the simulation sweep and the network training, and then applied automatically. The skill has not disappeared — it has been moved from the run to the setup.

Reproducibility improves as a side effect

Two analysts given the same spectrum can reasonably build two different models and reach two slightly different answers. That is uncomfortable in research and unacceptable in quality control. Removing human-defined initial parameters removes that variance, which for a manufacturer may matter more than the time saved.

Step in the workflowTraditional least-squares routeNetwork-assisted route
Choosing the structural modelAnalyst decides, per sampleGenerated by a trained network
Initial parameter valuesAnalyst supplies by judgementGenerated by a second network
Final fittingLeast-squares regressionLeast-squares regression, unchanged
Reported fit qualityBaselineSame quality of fit
Computational timeBaselineGenerally lower
Human time per sampleSubstantial and repeatedFront-loaded into training
Expertise required at run timeHighLow
Run-to-run reproducibilityDepends on the analystDeterministic given the same input

How it compares to the wider literature

Toledo’s group is not alone here. Published work on deep learning applied to ellipsometry has reported inference of complex refractive index and thickness roughly a thousand times faster than traditional nonlinear regression, framed explicitly as an enabler for high-throughput and autonomous materials development. The Toledo contribution is narrower and, arguably, more deployable: it keeps the regression and fixes the part that needed a person.

The Last Hurdle Is Real Cadmium Telluride Cells

Every simulation-trained model faces the same moment of truth, and Bordovalos named it himself.

The quote

“That’s the last hurdle,” he said. “I need to see if the model can make the jump handling the simulated stuff very well to handling the real stuff very well.” That is a clean statement of the simulation-to-reality gap, and it is the single most common failure mode in this class of work.

Why real ellipsometry data is harder

Simulated spectra are generated by the same physics the model is being asked to invert, so they are internally consistent by definition. Real ellipsometry spectra carry instrument response, alignment error, sample non-uniformity within the measurement spot, contamination, and structure the model was never told about. A network can score beautifully on held-out simulations and still fall over on the first plate off the line.

The convenient test bed

Toledo happens to be an unusually good place to run that test. Cadmium telluride solar cells are readily available at the university’s Wright Center for Photovoltaics Innovation and Commercialization, and thin-film photovoltaics is one of the group’s specialisms. The samples, the deposition equipment and the ellipsometry instruments are all in the same building.

What success would look like

The honest bar is not “the network gets the right number.” It is that structural decisions and parameter estimates from real measurements agree with what a skilled analyst would have produced, across enough samples and enough process conditions to be trustworthy. This is a transfer learning problem in everything but name, and the answer is empirical rather than theoretical.

What failure would look like

The realistic failure is not a spectacular one. It is that the ellipsometry tool works on the materials whose behaviour the simulation sweep anticipated, and quietly degrades on the ones it did not — a new dopant, an unexpected interfacial layer, a rougher surface than the sweep covered. That is a bounded and manageable outcome, and it is why the regression stays in the loop.

Who Alex Bordovalos Is and Why Toledo Matters

The person and the place are not incidental detail here. Both explain why this particular problem got picked up by this particular student.

Three degrees, one campus

Bordovalos earned a bachelor’s degree in physics from Toledo in 2020, then a Professional Science Master’s in Photovoltaics from the same university in 2022, and is now nearing the end of his doctorate. UToledo reports that he chose to stay — he is working toward his third degree at his hometown university.

The master’s programme is the mechanism

The Professional Science Master’s deliberately blends laboratory skills, business management and a real industry internship. That structure is what put a physics student inside a solar cell manufacturer at the exact moment he was capable of recognising an analysis bottleneck as a research problem rather than an annoyance.

The recognition so far

Beyond the journal article, Bordovalos has presented at the IEEE Photovoltaics Specialists Conference and the AVS International Symposium and Exhibition. The Spectroscopic Ellipsometry Technical Group gave him the J. A. Woollam Outstanding Student Poster Award at the latter in 2024 — a prize named after the company whose instruments made the technique commercially routine.

Toledo’s standing in thin films

Thin-film photovoltaics research at Toledo feeds a broader institutional strength in materials science, an area where the university’s researchers publish highly cited work and hold a global subject distinction in the U.S. News & World Report rankings. The Wright Center is the physical expression of that: a facility built around exactly the materials this tool needs to prove itself on.

YearMilestoneDetail
2020Bachelor’s degree in physicsUniversity of Toledo
2022Professional Science Master’s in PhotovoltaicsIncluded the industry internship that surfaced the problem
2024J. A. Woollam Outstanding Student Poster AwardAVS International Symposium and Exhibition
2025First-author paperJournal of Applied Physics, 3 September 2025
2026Real-sample testing beginsCadmium telluride cells at the Wright Center
Years elapsed since the 2020 bachelor’s degree
2022 — master’s completed 2 years
2024 — student poster award 4 years
2025 — first-author journal paper 5 years
2026 — testing on real cells 6 years

What Faster Ellipsometry Would Change on a Production Line

Podraza pointed at semiconductor and energy manufacturers. It is worth being concrete about what actually changes for them.

From ellipsometry audit to process control

The difference between a measurement you interpret in minutes and one you interpret in hours is the difference between adjusting a process and documenting it. Fast, unattended ellipsometry turns an ellipsometry instrument into a feedback sensor, and feedback sensors are what allow tighter tolerances without more scrap.

Full-map analysis becomes affordable

If every point on an 81-point map has to be fitted by hand, nobody fits every point. They fit three and hope. When the marginal cost of a fit collapses, uniformity across a whole plate becomes something you can actually see, which is precisely the non-uniformity problem that sent Bordovalos down this path in the first place.

Fewer specialists in the critical path

A plant that needs a PhD-level analyst on shift to release product has a staffing problem disguised as a technical one. Pushing the expertise into a trained model does not eliminate the need for expertise, but it stops that expertise being a per-sample dependency.

The sectors most exposed to slow ellipsometry

SectorWhat is measuredWhy analysis speed matters
Semiconductor fabricationGate oxides, barrier and dielectric layersInline process control at high wafer volumes
Thin-film photovoltaicsAbsorber, window and contact layersUniformity across large-area plates
Displays and OLEDOrganic stacks and encapsulationThin, multi-layer stacks with tight tolerances
Optical coatingsAnti-reflective and filter stacksMany layers, each one compounding error
Research laboratoriesNovel and unknown materialsNo established model to start from

Where Ellipsometry Automation Fits the Wider AI-for-Science Shift

This is one project at one university, but it is a very clear instance of a pattern that is now everywhere in instrumentation.

The pattern: invert the cheap direction

Wherever forward simulation is cheap and inversion is expensive, the same recipe applies. Simulate exhaustively, train a network on the synthetic pairs, and use the network’s output to initialise the rigorous method rather than to replace it. The same shape of solution shows up in crystallography, in scattering, in spectroscopy and now in ellipsometry.

Physics-trained people are building the tools

Almost one in five physics doctorate-earners entering the workforce in 2024 went into AI development, and roughly half of recent physics doctorate holders in permanent roles who use these tools are building machine learning models with them. Bordovalos is not an outlier; he is the modal case, and it is worth noticing that the discipline is producing its own tooling rather than waiting for a vendor.

Domain knowledge is the moat

The reason this works is that the simulation sweep encodes real physics. Anyone can train a network. Knowing that a film on soda-lime glass may need an optically distinct surface layer, and building that possibility into the training set, is domain knowledge — and it is what separates a useful scientific tool from a curve fitter. The same is true in every applied data science project we see: the modelling is rarely the hard part.

Where the AI tools stop

None of this is generative, none of it is conversational, and none of it is an agent. It is supervised learning against physically generated labels, with a deterministic solver downstream to catch the errors. That is a much older and much more reliable pattern than the one dominating the headlines, and for measurement work it is the right one.

What This Means If You Buy or Build Measurement Software

Most readers will never touch an ellipsometer. The transferable part is the procurement question underneath.

Ask where the training data came from

If a vendor claims automated interpretation, the first question is whether the model was trained on simulated data, measured data, or both — and what range the simulation covered. A sweep that only spans a narrow band around nominal will look excellent in a demonstration and fail on the day a process drifts.

Ask what happens when the model is wrong

The Toledo design is instructive because the network never has the last word. Its output initialises a regression that can be inspected and judged by its fit quality. Any measurement tool that cannot show its working, or that offers a number with no residual to check it against, is asking for trust it has not earned.

Ask whether the physics is in the loop

There is a meaningful gap between a model that has learned a physical relationship from physically generated examples and one that has learned correlations from a historical database. Both can be accurate. Only one degrades predictably. For anything feeding ML model development in a regulated or safety-relevant setting, that distinction is the whole ballgame.

Ask who owns the model when the student graduates

This is a doctoral project reaching its final year. Research code becomes abandonware with depressing regularity, and a tool that works brilliantly in one lab is not a product. If this matures into something a manufacturer can buy, it will be because somebody deliberately took on maintenance, validation and support — none of which is a physics problem.

Frequently Asked Questions About Ellipsometry

What is ellipsometry used for?

It measures the thickness and optical properties of extremely thin films — typically from under a nanometre to a few micrometres — by detecting how a sample changes the polarisation state of reflected light. It is standard practice in semiconductor fabrication, photovoltaics, display manufacturing and optical coatings.

Why is ellipsometry analysis so slow?

Because ellipsometry is indirect. The instrument returns a polarisation signature, not a thickness, so an analyst must build a layer model of the sample and adjust it until the predicted spectrum matches. Choosing that model and its starting values is judgement work, and it has to be repeated for every measurement point.

Does the neural network replace the physicist?

No. In the published Toledo method the networks generate the structural model and the initial parameter values, and a conventional least-squares regression still produces the final fit. The physicist’s judgement moves from the individual run into the design of the training sweep.

How much training data is involved?

Bordovalos says he is working with about half a million simulations, generated by varying expected parameters — thickness, for instance, by roughly 50% in either direction — across the range of structures the ellipsometry tool needs to recognise.

Has it been tested on real production samples?

Not yet, in the doctoral form described. The published validation used three laboratory-grown amorphous silicon films. Real-world testing on cadmium telluride solar cells at the Wright Center is the next step, and Bordovalos describes it as the last hurdle.

Is this available commercially?

No commercial release has been announced. The peer-reviewed ellipsometry method is published and reproducible, and the instrument vendors in this market have been moving toward automated analysis independently, but there is no product to buy off the back of this specific work.

What is spectroscopic about spectroscopic ellipsometry?

The measurement is repeated across many wavelengths rather than at a single one. That extra dimension is what makes it possible to separate thickness from optical constants — and it is also what makes each measurement a larger inverse problem to solve.

References