At 7nm and below, the feature geometries in production wafer layers are substantially smaller than the wavelength of visible light used in broadband brightfield optical inspection. A 7nm half-pitch metal layer uses 193i (immersion ArF) lithography with wavelength 193 nm; a broadband inspection system using wavelengths in the 250–800 nm range cannot resolve individual lines and spaces in these layers optically. This creates a classification problem that cannot be solved by increasing optical numerical aperture (NA) alone.
Yet optical inspection remains essential at advanced nodes — the throughput economics of e-beam inspection make it impractical as the primary screening method for high-volume wafer flows. The way inspection teams navigate this tension determines how much useful process information they extract from optical inspection data, versus how much is simply noise.
The Resolution Floor and What It Means for Classification
Optical resolution is governed by the Rayleigh criterion: R = 0.61λ/NA. For a broadband inspection system with effective wavelength around 400 nm and NA of 0.9, the theoretical resolution floor is approximately 270 nm. In practice, aberrations and coherence effects push the practical resolution somewhat higher. The structures you are inspecting in a 7nm metal layer are 20–40 nm wide.
This means that individual patterned features in advanced-node layers are below the optical resolution floor — they are not resolved as distinct structures in the inspection image. What is detectable is optical interaction effects: interference from sub-resolution gratings, scatter signatures from particulate contamination, and anomalous reflectance changes from localized layer thickness variations. Detection is real; classification based on resolved morphology is not possible.
Defect classification in the traditional sense — "this blob has the shape of a bridge short, that blob has the shape of a pit" — does not work at 7nm for in-layer defects. The image does not contain the morphological information. What the image contains is an optical signature that is correlated with the type of anomaly, but the correlation is indirect and requires a different classification approach.
Feature-Based Classification vs. Morphological Classification
Traditional ADC (automatic defect classification) systems used by mature-node inspection relied heavily on morphological features: defect area, aspect ratio, perimeter-to-area ratio, convexity, and gray-level statistics within the defect bounding box. These features encode shape information. Shape information is meaningful when the defect is resolved.
For advanced-node optical inspection, effective classifiers need to operate on features that capture the optical signature rather than the morphology. This includes: the spatial frequency content of the defect image patch, the relationship between brightfield and darkfield channel responses for the same location, the spectral response across illumination wavelength bands (if the system supports multi-spectral acquisition), and the local context within the die layout (what patterned structures surround the detection location).
The layout context point is often underappreciated. A particle on a dense metal fill region produces a different optical signature than the same particle on an open dielectric field, not because the particle is different but because the surrounding pattern modifies how light scatters and reflects from the region. A classifier that incorporates layout context as a feature — by cross-referencing detection coordinates against the design database — achieves substantially higher classification accuracy than one that treats the detection patch in isolation.
The Nuisance Kill Problem at Advanced Nodes
Nuisance kills — detections that are algorithmic artifacts of the patterned structure rather than real defects — increase dramatically at advanced nodes. In a 7nm metal layer, the dense repeating structure of the design creates a quasi-periodic optical background that contains spatial frequencies the inspection algorithm interprets as anomalies. Managing nuisance kill rate at these nodes requires explicit die-to-database comparison at detection locations, not just die-to-die comparison.
Die-to-die comparison (the traditional approach) compares two adjacent die and reports differences. At advanced nodes, systematic patterning variations across the exposure field — EUV stochastic effects, overlay errors that shift periodically — create systematic differences between adjacent die that are not defects. Die-to-database comparison compares the observed optical image against a rendered reference image generated from the GDS design database at the same coordinates. Systematic layout differences appear in both the observed image and the reference, so they do not generate detections. Only genuine process anomalies — contamination, local etch variation, CDU outliers — appear as differences.
We are not saying die-to-die comparison is wrong at advanced nodes — for macroscopic defects and contamination events it remains effective and fast. The point is that for classification-quality signal at 7nm and below, die-to-database comparison at candidate detection sites is a prerequisite, not a luxury.
Combining Optical Detection with E-Beam Review
The practical architecture for 7nm inspection combines optical screening (high throughput, captures particles and macroscopic yield killers) with targeted e-beam review at candidate sites identified by optical detection. The optical system generates a candidate list — coordinates of potential defects with low-confidence optical classification. The e-beam review tool (SEM-based review) examines these candidates at nanometer resolution and provides ground-truth morphological classification.
The value of the optical system in this architecture is not standalone classification — it is accurate defect coordinate generation with low enough false discovery rate that the e-beam review queue remains manageable. An optical system generating 50 candidate sites per wafer for e-beam review is useful. One generating 2,000 candidate sites per wafer means the e-beam review tool is the throughput bottleneck, and the optical screening step has added cost without adding usable process signal.
Getting the optical-to-e-beam candidate rate right requires careful nuisance kill management at the optical step. The target: identify all wafers with real yield-relevant defects (sensitivity), while keeping the per-wafer candidate count low enough that e-beam review can keep pace with production flow (nuisance control). These two objectives trade off, and the operating point should be set based on production volume and available e-beam capacity — not as a fixed spec.
Classifier Training Data at Advanced Nodes
Any classification model needs training data. At advanced nodes, labeled training data for optical inspection classifiers is scarce for two reasons: production experience at these nodes is newer and therefore history is shorter, and the e-beam review confirmation loop required to generate labeled data for sub-resolution optical detections is expensive.
Active learning approaches — where the classification model selects its own most uncertain detections for e-beam review label acquisition — reduce the e-beam review cost per labeled sample compared to random sampling. Starting from an optical candidate list, the model flags the 5–10% of candidates with lowest classification confidence; e-beam review prioritizes those; new labels are added to the training set. The model improves faster per unit of e-beam time than random labeling would achieve.
Transfer learning from mature-node classifiers to advanced-node applications accelerates initial deployment when ground-truth data at the target node is sparse. The pretrained features from a mature-node optical classifier are not directly applicable (the optical signatures are different), but the feature extraction layers trained on large mature-node datasets provide better initialization than random weights. Node-specific fine-tuning on limited advanced-node data outperforms training from scratch on the same limited data.
What Process Engineers Should Expect
A realistic expectation for optical inspection classification performance at 7nm: high sensitivity to particle contamination events (particles ≥ 150 nm show reliably detectable optical signatures), moderate sensitivity to pattern defects (bridging shorts, opens) that extend over multiple sub-resolution features, and limited ability to classify the nature of subtle within-pattern anomalies without e-beam confirmation. The optical system identifies where to look; e-beam confirms what it is.
Process engineers who expect standalone optical classification accuracy at advanced nodes comparable to what they achieved at 28nm will be disappointed. Those who design their inspection workflow around optical screening plus targeted e-beam review, with classifier models that incorporate layout context, will extract genuinely useful yield-relevant signal from their optical inspection investment.