In-situ ellipsometry watches a surface while deposition, etching, annealing, adsorption, oxidation, or another process is still occurring. That timing changes the measurement from a final-state thickness check into a record of how the state evolved. A useful implementation therefore joins an optical instrument, a reactor-compatible line of sight, a time-dependent optical model, and process context such as valve states, plasma power, temperature, pressure, and cycle number. The result can reveal incubation, nucleation, coalescence, steady growth, composition drift, roughening, and endpoint—but only when the analysis separates genuine sample evolution from temperature, gas, window, and alignment effects.

In-situ, real-time, inline, and ex-situ describe different measurement relationships. In-situ means the wafer or witness sample is measured in the process environment without removal. Real-time means acquisition and analysis are fast enough to resolve the event of interest; an in-situ spectrum collected once per minute is not real-time for a two-second plasma transient. Inline metrology measures a wafer between process steps on the manufacturing line, while ex-situ measurement occurs after removal to a separate instrument. These categories may overlap, but they are not interchangeable.

Ellipsometry measures a polarization ratio rather than an absolute intensity alone. For an isotropic, nondepolarizing sample with no p–s coupling, the complex ratio is

The time coordinate is as important as wavelength. Each acquired spectrum is a slice through a trajectory whose latent state may include thickness, complex dielectric function, void fraction, roughness, interface width, or composition. A single-wavelength instrument can acquire quickly but may confuse thickness with optical constants. Spectroscopic acquisition adds independent information and makes model failure more visible, although the spectrum must still be fast compared with the process dynamics.

Time resolution must be designed around the fastest physical claim. The effective cadence includes exposure, polarization-state sequencing, spectral readout, transfer, fitting, and any averaging—not merely the detector integration time. Sequentially acquired polarization states can encode temporal evolution as false polarization when the sample changes during a cycle. Fast rotating-element, multichannel, or snapshot architectures reduce that risk, while slower full-spectrum scans may be appropriate for steady CVD growth. Report timestamps at the sample, dropped frames, synchronization latency, and the temporal impulse response used by filtering.

A deposition rate derived from fitted thickness is a time derivative, so noise amplification and model correlation matter. For discrete measurements, a local slope over a declared interval is more defensible than point-to-point differencing:

The first expression is the least-squares thickness slope in a time window; the second is average growth per cycle over a stated cycle interval. Neither proves a microscopic mechanism by itself. A changing fitted thickness can absorb changes in density, composition, surface coverage, temperature-dependent optical constants, or roughness when the model lacks those degrees of freedom.

A dynamic optical model must distinguish evolving state from fixed structure. Start with the known substrate and stable underlayers, then identify which parameters can physically vary during the observation window. A growing transparent layer may permit thickness to vary while its dispersion remains fixed. Nucleating metals often require a changing effective dielectric function or physically motivated island-to-coalescence model because a bulk-metal optical constant is invalid below percolation. Reactive films may need composition-dependent optical functions; rough or porous layers may use an effective-medium approximation only when its assumptions and parameter correlations are tested.

Temperature belongs in the model. Heating changes substrate and film optical constants, expands layers, bends fixtures, and modifies viewport stress birefringence. Gas admission changes refractive index along the beam path, and plasma emission can contaminate detected intensity or saturate a spectrometer. A baseline acquired at room temperature in vacuum is not automatically valid at process temperature and pressure. Whenever possible, separate a thermal soak, gas-only sequence, plasma-only blank, and actual surface reaction so their signatures can be compared.

Cycle-resolved analysis exposes kinetics that a final thickness cannot reveal. In ALD, align optical frames with precursor, purge, co-reactant, and plasma events. Compare equivalent phases of successive cycles rather than treating gas-phase transients as deposited material. Incubation appears as a nonsteady early trajectory; island nucleation and coalescence can change roughness and dielectric response before a stable linear thickness regime emerges. Dose saturation is demonstrated by a plateau versus exposure under adequate purge, not by a single apparently constant growth-per-cycle number. In ALE, a periodic optical trajectory may follow modification and removal half-cycles, but net thickness loss must be distinguished from reversible surface chemistry.

For CVD, PVD, and epitaxy, the time trace can separate nucleation delay from steady-state rate and identify transitions between amorphous, crystalline, porous, or coalesced regimes. Optical libraries measured at the growth temperature can connect dielectric response with composition, but library interpolation should carry calibration uncertainty. For reactive sputtering, a change in optical constants may track composition while chamber hysteresis changes the deposition state. Independent X-ray, electron, electrical, mass, or compositional measurements are valuable because ellipsometry alone does not chemically identify every change.

Etch monitoring requires a model that can survive disappearing layers. A robust model changes topology when a film reaches zero thickness, rather than permitting a negative thickness or forcing the residual into roughness. Endpoint can be defined from fitted remaining thickness, a stable change in derivative, a spectral feature, or a residual-tested state transition. Patterned product wafers may produce diffraction, depolarization, or lateral averaging that a planar witness model cannot represent; a chamber witness can track chemistry without necessarily reproducing feature-scale loading, selectivity, or aspect-ratio-dependent etching.

A statistical endpoint rule should include persistence and uncertainty. One example declares completion only when the upper confidence bound on remaining thickness stays below a threshold for several frames:

Here the threshold, confidence multiplier, and persistence count are recipe-specific engineering choices. The uncertainty must include model inadequacy and correlated drift where relevant; a covariance matrix from an overconfident least-squares fit is not a complete error budget. Before automatic control, replay historical traces, run the estimator in shadow mode, define fail-safe behavior for missing spectra or rising residuals, and bound the controller’s authority.

Viewport behavior is part of the measurement system, not background housekeeping. Deposited films on windows alter transmission, polarization, and angle-dependent phase. Etch products can haze a window; heating can create birefringence; purge flow can change gradients; and mechanical motion can alter the beam footprint. Use shutters, purge curtains, replaceable coupons, witness channels, reference beams, or periodic baseline checks as the reactor permits. A scalar intensity normalization cannot generally correct a polarization-changing window. Calibration should represent the installed optical train and be repeated after window replacement, cleaning, source or detector service, or alignment change.

Residuals provide an early warning. Plot residual versus wavelength and time rather than only a scalar mean-squared error. A smooth spectral bias can indicate a missing layer or incorrect dispersion; a sudden broadband offset can mark a window, plasma-emission, or alignment event; periodic residuals synchronized to valves can expose gas-path effects. Track parameter covariance and boundary hits. When thickness and refractive index become strongly correlated, freeze one only with external evidence or redesign the measurement with more wavelengths, a second angle, or an independent sensor.

Define process event and required time resolution
  -> Select wavelength range, incidence geometry, and reactor ports
  -> Calibrate installed optical train at relevant temperature and pressure
  -> Run thermal, gas, plasma, and blank-substrate controls
  -> Build the simplest physically valid dynamic layer model
  -> Synchronize spectra with recipe telemetry and a common clock
  -> Fit time series with continuity constraints and inspect spectral residuals
  -> Separate incubation, transient, steady, and endpoint regimes
  -> Validate thickness, composition, or structure independently
  -> Run endpoint or controller in shadow mode with fail-safe rules
  -> Release bounded control and monitor drift, windows, and uncertainty

Process control needs observability, latency discipline, and explicit failure states. Observability asks whether the available spectra uniquely constrain the state needed for control. Latency includes acquisition, fitting, decision, communications, and actuator response. Failure states include beam loss, saturated detector, plasma flash, invalid model topology, stale recipe telemetry, a parameter at its bound, or residuals outside the validated domain. A safe controller rejects low-confidence estimates, holds or terminates according to the recipe’s hazard analysis, and records why it acted.

Store raw detector data where practical, Ψ and Δ spectra, polarization conventions, timestamps, recipe event logs, chamber conditions, calibration versions, window history, model topology, optical-function provenance, parameter bounds, residual spectra, covariance, filter settings, and software version. Preserve the unfiltered data even when a smoothed trend drives the display. Report whether thickness is physical, effective-medium, or optical-equivalent and whether rate is instantaneous, windowed, cycle-averaged, or wafer-averaged.

Qualification should challenge the complete chain. Use known films spanning the working thickness and optical-constant range, temperature-matched controls, deliberate window contamination limits, acquisition dropouts, endpoint overshoot tests, and repeat runs across maintenance states. Compare against ex-situ ellipsometry only with a compatible model and account for oxidation, moisture uptake, relaxation, or temperature change after removal. Agreement with profilometry, X-ray reflectivity, microscopy, mass sensing, or composition analysis strengthens the physical interpretation when those methods probe the same region and state.

The strongest in-situ conclusion is a synchronized trajectory with falsifiable controls. Ellipsometry detects optical evolution with exceptional sensitivity, but kinetics emerge only after the model, timing, environmental baselines, and independent checks support the assignment. A precise-looking curve without those controls may be a record of a window or an assumed dielectric function rather than the wafer process.

The durable way to interpret in-situ ellipsometry is through a polarization-trajectory-time-resolution-dynamic-model-process-synchronization-window-drift-endpoint-confidence-and-independent-validation lens.