How to Learn Spectroscopic Ellipsometry and Thin-Film Optical Metrology: From Polarization Changes to Thickness, Optical Constants and In-Situ Growth

Wait, What? Ellipsometry Usually Does Not Measure Film Thickness Directly

An ellipsometer does not point a ruler at a film. It sends polarized light toward the sample, measures how reflection changes the polarization state, and from that change—plus an optical model—infers thickness, refractive index n, extinction coefficient k, roughness and sometimes composition.

ellipsometry is an inverse problem: the instrument measures polarization change; thickness appears only after a model earns the right to call it thickness.

The One-Sentence Answer

Learn spectroscopic ellipsometry by tracing s/p-polarized reflection → complex Fresnel ratio → measured Ψ and Δ → multilayer optical model → fitted n, k and thickness, then add roughness, anisotropy, backside reflections, parameter correlation and in-situ growth while testing whether alternative models fit equally well.

Beginner Layer — Polarization as the Receiver

Stage 1: Light Has an Electric-Field Direction

Polarization describes the orientation and evolution of the transverse electric field.

Stage 2: Reflection Treats s and p Polarization Differently

Relative to the plane of incidence, s is perpendicular and p parallel; their reflection coefficients are generally different.

Stage 3: Ellipsometry Measures Their Ratio

ρ = rp/rs = tan(Ψ)e^(iΔ). The primary observables are Ψ and Δ.

Stage 4: A Film Changes Polarization Through Interference

Reflections from the air/film and film/substrate interfaces interfere. Film thickness and refractive index set their relative phase.

Optical-Constant Layer

Stage 5: Refractive Index Can Be Complex

ñ = n + ik, where n controls phase and k absorption.

Stage 6: Dielectric Function Is Equivalent Information

ε = ε₁ + iε₂ is related to n and k.

Stage 7: Optical Constants Depend on Wavelength

Real materials are dispersive; spectroscopic ellipsometry gains power by fitting many wavelengths simultaneously.

Thickness Layer

Stage 8: Thin Transparent Films Produce Interference Structure

Optical path length depends on thickness, refractive index and incidence angle.

Stage 9: Thickness and n Can Trade Off

Different thickness/n combinations can generate similar phase shifts, making apparently precise fits non-unique.

Stage 10: Multiple Angles Add Constraints

A model that fits one incidence angle but fails another is incomplete.

Transparent-Film Models

Stage 11: Cauchy Dispersion Is Useful in Transparent Regions

n(λ) = A + B/λ² + C/λ⁴ + … is useful where absorption is negligible.

Stage 12: Too Many Cauchy Terms Can Overfit

The simplest physically justified dispersion model is usually stronger than a numerically flexible one.

Absorbing-Film Models

Stage 13: Absorbing Materials Need Physically Constrained Dispersion

Lorentz, Tauc–Lorentz, Cody–Lorentz and Drude terms model different electronic responses.

Stage 14: Drude Terms Describe Free-Carrier Response

Conducting materials can show strong long-wavelength carrier response related to plasma frequency and scattering rate.

Surface Roughness

Stage 15: Roughness Changes the Effective Optical Boundary

Ellipsometry often represents roughness as an effective-medium layer.

Stage 16: Effective-Medium Roughness Is a Model, Not an AFM Scan

A fitted EMA thickness need not equal AFM RMS roughness.

Stage 17: Roughness and Composition Gradients Can Mimic Each Other

Physical roughness, porosity and graded composition can produce similar Ψ/Δ changes; AFM/TEM/XRR help constrain them.

Multilayer Systems

Stage 18: Each Additional Layer Adds Parameters

Substrate, native oxide, target film, roughness and contamination can all contribute.

Stage 19: A Wrong Buried Layer Can Bias Everything Above It

If substrate or native-oxide properties are wrong, the target layer can compensate numerically.

Stage 20: Parameter Covariance Should Be Inspected

Low MSE does not reveal strong parameter correlation.

Backside Reflection

Stage 21: Transparent Substrates Can Reflect From the Rear Surface

That extra beam can alter measured polarization.

Stage 22: Backside Reflection Can Masquerade as Film Structure

Roughening/blackening the backside, wedged substrates or incoherent-backside models can suppress or model the effect.

Stage 23: ISO 23131-3:2026 Makes These Limits Explicit

The 2026 standard addresses backside reflection, anisotropy and sample alignment in model-based spectroscopic ellipsometry.

Anisotropy Layer

Stage 24: Some Films Have Direction-Dependent Optical Constants

Crystals, stretched polymers and aligned nanostructures can be birefringent.

Stage 25: Generalised Ellipsometry Handles Polarization Coupling

Mueller- or Jones-matrix approaches are needed when s and p mix.

Stage 26: Depolarisation Is Information

Rough, patterned or laterally nonuniform samples can depolarize light; averaging should not hide this.

Thin and Ultrathin Films

Stage 27: Ultrathin Films Create Strong Parameter Correlation

Monolayer-scale signals depend strongly on substrate, oxide, roughness and assumed dispersion.

Stage 28: Thickness-Only Fits Can Be Robust When n,k Are Independently Known

Reducing free parameters makes the inverse problem easier to identify.

Semiconductor Application Layer

Stage 29: Ellipsometry Is a Workhorse for Semiconductor Films

It measures oxides, photoresists, dielectric layers, epitaxial films and composition-dependent optical response without contact.

Stage 30: Band-Gap Models Need Care

Optical gap estimates depend on model, disorder, excitons, thickness and spectral range and should not automatically be equated with an electrical transport gap.

In-Situ Growth

Stage 31: Ellipsometry Can Watch a Film Grow in Real Time

Optical access enables Ψ/Δ monitoring during ALD, CVD and epitaxy.

Stage 32: In-Situ Data Can Reveal Nucleation Delay

Incubation, island coalescence and density changes can precede steady growth.

Stage 33: Dynamic Models Can Track Evolving Optical Constants

If n and k change during crystallisation or densification, a fixed-optical-constant model can put material-state change into apparent thickness.

Infrared and Advanced Ellipsometry

Stage 34: Infrared Ellipsometry Probes Phonons and Free Carriers

IR ellipsometry extends to lattice vibrations, carrier absorption and narrow-gap semiconductors.

Stage 35: THz Ellipsometry Extends to Lower Energy

Longer wavelengths probe conductive sheets, anisotropy and carrier dynamics.

Model Validation

Stage 36: MSE Is Not Physical Truth

A low MSE only means the model reproduces measured Ψ/Δ numerically.

Stage 37: Residual Structure Is Diagnostic

Systematic residuals versus wavelength or angle can reveal a missing oscillator, backside reflection, anisotropy or wrong layer.

Stage 38: Fit One Physical Regime at a Time

Determine substrate, fit transparent region, add absorption physics, test multiple angles and compare alternative layer models.

Hybrid Metrology

Stage 39: Ellipsometry + AFM Helps Separate Thickness From Roughness

Topography and optical response constrain different parts of the inverse problem.

Stage 40: Ellipsometry + TEM/XRR Provides Orthogonal Thickness

Agreement across independent receivers strengthens metrology.

AI and Inverse-Modelling Frontier

Stage 41: Machine Learning Can Accelerate Inversion

Neural networks can estimate thickness, n/k parameters and material classes rapidly.

Stage 42: ML Learns the Model Family It Was Trained On

A network trained only on smooth isotropic films can fail confidently on graded, anisotropic or backside-reflecting samples.

Stage 43: Physics-Informed Models Preserve Fresnel Constraints

Hybrid approaches can use optical forward models with ML initialization, acceleration or anomaly detection.

Stage 44: Professional Ellipsometry Is an Inverse-Model-Identifiability Problem

Which thickness and optical constants are genuinely constrained by Ψ/Δ after substrate response, backside reflection, roughness, anisotropy, parameter covariance and dispersion-model choices are all allowed to compete?

Evidence: What Makes an Ellipsometry Result Strong?

Stronger evidence combines multiple incidence angles, calibrated wavelength/angle, substrate measurement, physically appropriate dispersion, residual inspection, covariance analysis, alternative-model comparison and AFM/TEM/XRR constraints.

Misconceptions Worth Hunting

Ellipsometry directly measures thickness.
Ψ and Δ are material constants.
A low MSE proves the layer model.
Cauchy fits are valid in absorbing regions.
Effective-medium roughness equals AFM roughness exactly.
Backside reflection is negligible for all transparent substrates.
More fitted oscillators always improve physical accuracy.
Isotropic models are safe for every thin film.
An optical band gap equals an electrical gap automatically.
ML eliminates non-uniqueness.

Transfer Check

A transparent-film fit returns 100.0 nm, but a second model with slightly different n gives 92 nm with nearly identical residuals. Is thickness uniquely known? No. Thickness–n correlation is exposed.

A glass-substrate spectrum shows periodic oscillations that vanish after backside roughening. Were they film interference? Probably not; backside reflection was the stronger explanation.

AFM reports 3 nm RMS roughness while ellipsometry fits a 7 nm EMA roughness layer. Must one be wrong? No. The two roughness quantities are not identical physical definitions.

A neural network returns 25 nm for an anisotropic film although it was trained only on isotropic stacks. Is the number trustworthy? Not without out-of-domain validation.

How We Know the Learning Has Held

A learner should be able to explain s/p polarization, Ψ/Δ, interference-based thickness sensitivity, complex n/k, dispersion models, thickness–n correlation, roughness EMA, multilayers, backside reflection, anisotropy/generalized ellipsometry, in-situ growth, residual/covariance analysis and ML inverse-model limits.

Model Limits

Ellipsometry is extremely precise when the optical model is appropriate and can be precisely wrong when the layer stack, substrate, roughness or anisotropy model is wrong.

measured Ψ/Δ + angle/wavelength calibration + substrate + layer stack + dispersion model + roughness + backside + anisotropy + covariance + orthogonal metrology

Teaching Guide

polarization → s/p Fresnel reflection → Ψ/Δ → interference → n/k → thickness → multi-angle → Cauchy/absorbing models → roughness → multilayers → backside reflection → anisotropy → ultrathin films → in-situ growth → IR/THz → residuals/covariance → hybrid metrology → ML.

“If the instrument never touches the film and never directly reads a ruler, where does the reported thickness actually come from?”

Connect This to the eduKate Learning Estate

Research Foundations and Further Learning

Core ellipsometry literature defining ρ = rp/rs = tanΨ exp(iΔ).
ISO 23131-3:2026 — model-based spectroscopic ellipsometry for transparent single-layer films.
Current literature on Cauchy-model parameter correlation and overfitting.
Effective-medium roughness and graded-interface studies.
2026 review of infrared spectroscopic ellipsometry for narrow-band semiconductors and hybrid metrology.

The Quiet Ending

The beginner asks: “How thick is the film?”

The developing optical scientist asks: “What Ψ and Δ changed?”

The advanced learner asks: “Which combination of n, k, roughness and thickness can explain that change?”

Which film parameter remains identifiable after the optical model itself is treated as a hypothesis rather than a hidden truth machine?

Small Group Tutorials