Broadband Plasma Inspection
Broadband plasma defect inspection is the workhorse of unpatterned and patterned wafer inspection at advanced nodes. A laser-sustained plasma source can provide a broad UV-visible continuum; the delivered wavelength bands are tool- and configuration-specific, and that light is delivered to the wafer through a high numerical aperture objective, and scattered or reflected photons are collected and thresholded against a reference to find defects far smaller than the illuminating wavelength.
Why broadband rather than a single laser line
A monochromatic source produces thin-film interference that can render a defect invisible on one stack and obvious on another, because the reflected field depends sharply on wavelength and layer thickness. Spectral diversity can reduce sensitivity nulls and speckle relative to narrowband illumination, but it does not erase thin-film or coherent-pattern effects. Recipes still optimize wavelength, angle, polarization and pupil filtering for each layer.
Scattering physics sets the detection limit
The signal is scattered power, and its dependence on particle size $d$ changes regime as $d$ approaches the wavelength $\lambda$. The size parameter
separates the two cases. For $x \ll 1$ the particle is a dipole and Rayleigh scattering applies, with cross-section
where $m$ is the complex refractive index ratio of particle to medium. Two dependencies dominate inspection economics. The $d^{6}$ term means halving the defect size cuts signal by a factor of 64, so sensitivity falls off a cliff rather than degrading gracefully. The $\lambda^{-4}$ term is why inspection wavelengths keep moving deeper into the UV: dropping from 355 nm to 200 nm recovers roughly an order of magnitude in scattered power at fixed defect size.
Once $x \gtrsim 1$ the dipole approximation fails and the full Mie solution is required. Scattering becomes strongly forward-peaked and oscillatory in $d$, so signal no longer rises monotonically with size, and a sizing curve calibrated on polystyrene spheres mis-sizes real defects of different refractive index.
Broadband advantage comes from channel diversity, not from averaging every wavelength into one number. Each spectral, angular, and polarization channel samples a different complex interference condition; a defect near a null in one channel may have useful contrast in another. The recipe must weight channels using calibrated signal, noise, correlation, and review-confirmed capture while respecting photon dose and scan-time limits.
Brightfield and darkfield are different instruments in one tool
Brightfield collects the specular reflection, so a defect appears as a local decrease in an otherwise bright return. The background is large, which means shot noise is large, and sensitivity is limited by contrast against that background. It excels on defects that perturb reflectivity or phase, such as residues, voids and pattern deformation.
Darkfield rejects much of the specular beam and collects off-axis scattered light, so the background can be reduced and a small particle produces a bright point on a dark field. For some sub-wavelength particles on sufficiently smooth unpatterned surfaces, rejecting the specular return can improve signal-to-background enough to make darkfield highly effective. Performance still depends on haze, roughness, collection geometry, particle material and shape, photon statistics, and contamination risk. On patterned wafers the pattern itself scatters, and that diffracted light must be suppressed by Fourier-plane filtering before the defect signal is recoverable.
| Mode | Background | Best for | Limiting noise |
|---|---|---|---|
| Brightfield | high specular | residue, void, pattern defect | photon shot noise |
| Darkfield | reduced specular; residual haze/scatter remains | small particles, scratches | pattern scatter, surface haze |
The nuisance problem and learned filtering
A modern inspection pass on a production layer can return hundreds of thousands of events, of which only a small fraction are killer defects; the rest are nuisance — grain contrast, line-edge roughness, colour variation, benign residues. Classical thresholding cannot separate them because nuisance and killer populations overlap in brightness. Production flows therefore feed multi-channel event attributes into a learned classifier, trained against SEM-reviewed ground truth, which estimates
The operating point is chosen on the resulting precision-recall curve rather than by a fixed threshold, because the cost of a missed killer defect and the cost of a wasted SEM review are wildly asymmetric. Classifier benefit is layer-, population- and review-ground-truth-dependent. Its operating point must be monitored for drift and rare-defect recall; nuisance reduction without a confidence interval is not evidence of unchanged capture.
Where it sits against e-beam inspection
Broadband plasma trades absolute resolution for throughput. E-beam inspection can reveal smaller or electrically contrasting defects that a given optical recipe misses, but its area throughput and charging or dose constraints usually make it selective rather than a universal full-wafer substitute. The relative throughput depends on tool architecture, pixel size, dose, review strategy, and layer. Broadband optical inspection can cover large wafer areas at production-oriented cadence, which is what makes it the sampling backbone of excursion control: it is the tool that notices a chamber has drifted before a lot is lost.