Defect inspection and review workflow is the inline machinery that turns a raw pattern of light scattered off a wafer into an actual engineering decision about whether that wafer, and every wafer behind it in the lot, should keep moving through the fab. An optical inspection tool first finds candidate defect sites across the wafer using brightfield or darkfield imaging, a review step then revisits each of those coordinates at higher magnification to characterize what was actually found, a classifier bins each confirmed defect into a class, and a disposition decision, rework, accept, or contain, closes the loop before the wafer advances. None of those four steps works in isolation; the value of the whole workflow comes from how tightly the output of each stage feeds the next one, and from how consistently the accumulated defect data feeds back into a fab's broader yield learning system.

Inline optical inspection using brightfield or darkfield imaging is the workflow's entry point, scanning either the full wafer or a sampled subset of die to flag coordinates where the reflected or scattered light pattern deviates from an expected reference. Brightfield inspection illuminates the wafer directly and is generally more sensitive to larger, higher-contrast defects such as residue or scratches, while darkfield inspection collects only scattered light and tends to pick up smaller particles and subtle pattern anomalies that brightfield imaging can miss entirely. A routine monitoring recipe commonly samples 20% to 40% of total die area rather than the full wafer, trading some detection completeness for the throughput needed to keep inspection paced with the production line, while an excursion investigation typically reverts to full-wafer, full-density sampling until the root cause is confirmed. Critical layers, where a small defect has an outsized yield impact, are routinely sampled at two to three times the density used for a non-critical layer.

Defect review takes each flagged coordinate from inspection and revisits it under a higher-resolution imaging tool, typically a review SEM, to confirm the defect is real and to capture the image detail a classifier needs. Because an optical inspection tool trades resolution for throughput, a meaningful fraction of flagged coordinates, often in the range of 10% to 20% depending on recipe sensitivity, turn out on review to be nuisance signals such as grain structure or a stage-positioning artifact rather than genuine defects, and filtering those out before classification keeps the downstream Pareto data meaningful. Review SEM imaging routinely resolves feature detail below 50 nm, fine enough to distinguish a genuine pattern defect from a similarly sized particle that optical inspection alone could never separate reliably. A well-tuned review recipe balances magnification and field of view carefully, since too tight a field of view risks missing the defect entirely if stage-to-stage coordinate accuracy drifts by even a couple of µm. Review throughput is a real constraint on the whole workflow, since a single high-magnification image can take on the order of 1 s to 2 s to capture and store, and a lot with several hundred flagged coordinates can therefore consume a meaningful fraction of tester and operator time before classification even begins.

Classification bins each confirmed defect into a class such as particle, scratch, pattern defect, or residue, and that binned data is what turns individual defect counts into a ranked Pareto a team can act on. A mature classification recipe on a stable layer typically holds accuracy in the 85% to 95% range, with lower-confidence calls routed to a human reviewer rather than committed automatically, and the resulting Pareto chart, ranking defect classes by count, routinely shows the top two or three classes accounting for 70% to 80% of the total flagged population on a given lot. A shift in Pareto rank order from lot to lot is treated as seriously as a change in total defect count, since a normally minor class suddenly climbing the ranking often points more directly at which specific process module just changed than the raw count trend does on its own.

Disposition is the decision point where classified defect data becomes an action: accept the lot as-is, rework it if the process allows, or contain it for engineering hold pending further analysis. A typical flagged lot sees roughly 70% to 85% of cases dispositioned as accept, since most flagged defects, once classified, fall within an established risk tolerance for that defect class and density, while 10% to 20% go to rework when the process step allows a corrective action such as a repeat clean or strip-and-redo. The remaining share, usually below 10% of flagged lots, is contained and held for engineering investigation, a disposition reserved for cases where defect density, class, or spatial pattern suggests a genuine yield risk rather than routine background noise. Disposition rules are typically encoded so that a spatial cluster of otherwise unremarkable defects, several sites close enough together to suggest one root cause, escalates a lot to contain status even when the total defect count alone would not have triggered a hold.

Every defect record generated across inspection, review, and classification flows into a yield management system that aggregates data across lots, layers, and tools, turning individual wafer events into a fab-wide learning signal. Each record is tagged with wafer ID, die and field coordinate, defect class, and process layer, so that a yield engineer can later query the accumulated dataset for a specific tool's defect signature or a specific layer's historical Pareto trend rather than working from a single lot in isolation. Statistical process control rules built on top of that aggregated data trigger an automated hold when a defect class exceeds two to three times its rolling baseline count, or when overall defect density on a layer rises by 30% or more relative to the prior several lots. Correlation studies tying inline defect density to final die sort yield routinely show that lots flagged with an above-baseline defect count see a yield penalty of several percentage points relative to lots that pass inspection clean, which is exactly the evidence that keeps engineering leadership funding the inline inspection and review infrastructure rather than treating it as pure overhead.

Escalated or ambiguous classification calls are routinely confirmed by physical failure analysis before a disposition decision is finalized on a high-stakes lot, closing the loop between an automated call and a verified root cause. AFM topography resolves surface height differences fine enough to distinguish a genuine pit from a shallow residue patch that looks similar in a plan-view review image, while SIMS depth profiling and XPS surface analysis identify the chemical composition of a suspected contamination-class defect and trace it back to a specific upstream chemistry. DLTS spectroscopy is occasionally brought in when a pattern defect is suspected of introducing an electrically active trap level, tying a purely visual defect call to a measurable device-level consequence before a large volume of product is contained on the strength of an inspection image alone. Some fabs additionally pull four-point probe sheet-resistance readings from the same lot to check whether a spatial defect cluster lines up with an electrical resistivity anomaly, since a defect pattern that correlates with a parametric signature is treated as far more likely to be yield-relevant than one that shows no electrical footprint at all.

Optical inspection flags candidate sites → Review SEM confirms and images each site → Classifier bins confirmed defects by class → Build ranked defect Pareto for the lot → Disposition decision: accept, rework, or contain → Route defect records to yield management system → SPC rules flag rank shifts and density spikes → Escalate to AFM, SIMS, XPS, or DLTS failure analysis

Viewed through an inline-defect-to-yield learning lens, the defect inspection and review workflow earns its place as one of the fab's most heavily instrumented processes because it converts a raw scattering pattern on a wafer surface into a disciplined chain of confirmation, classification, and disposition decisions, each one traceable back into a yield management system that turns individual inspection events into the fab-wide learning that keeps yield improving lot over lot.