Brightfield Inspection

Brightfield inspection finds wafer anomalies by imaging reflected light and looking for departures from the expected pattern. The useful output is a set of candidate defect locations, not a verdict that every bright or dark spot will kill a chip. Pattern edges, films, particles, and process variation all change the optical signal. A good recipe makes the defects of interest stand out while keeping ordinary variation manageable.

Brightfield keeps the reflected background in the image. In reflected-light brightfield, the collection path includes the specular return, the mirror-like reflection from the surface. Defects perturb that background through changes in reflectivity, scattering, absorption, or interference. Darkfield instead suppresses the specular return and collects selected scattered light. These are complementary ways of observing the same wafer, not a universal division into “pattern defects” versus “particles.”

The comparison must be between equivalent locations. A die-to-die method compares corresponding pattern regions on different dies; a cell-to-cell method can exploit repeated structures within a die. Images must be registered so normal edges line up. Brightness normalization and treatment of legitimate process variation also matter. Without these steps, a small offset can create paired bright and dark residuals along many edges and overwhelm the real anomalies.

The simple difference image below explains the idea, not a proprietary detection algorithm. Two matching lines disappear on subtraction; an added bridge leaves a localized residual. Real inspection uses more sophisticated references, filtering, noise estimates, and classification. If the same error appears identically in both test and reference regions, a simple pairwise subtraction can cancel it too. Reference selection and complementary inspection are therefore part of coverage, not housekeeping.

Detection is not the same as resolving the shape. A small defect can change the measured signal even when the optics cannot reproduce its outline as a sharp image. Whether that change is detectable depends on contrast, photon statistics, background variation, collection optics, sampling, and the algorithm. Pixel size alone is not a minimum detectable defect diameter. Conversely, drawing more pixels across a blurred image does not recover information the optics never collected.

The film stack changes what is visible. A bridge on one layer, a residue on another, and a particle on a third can have very different contrast. Illumination spectrum, polarization, focus, and collection settings interact with the materials and patterned background. A shorter wavelength is not automatically the best setting for every defect. Choose optical conditions using representative defects on the actual layer, then validate them against normal process variation. Broadband and laser-based systems provide different implementations; neither label alone specifies sensitivity.

A threshold trades capture against workload. In a simplified detector, a candidate is reported when its difference score exceeds a threshold. Lowering that threshold can recover weak defects, but also admits nuisance signals. Raising it can make the map look cleaner while hiding relevant defects. Real tools may use spatially varying thresholds and multiple classifiers rather than one global number, but the qualification question remains: which important defects are captured, and how much review effort does that capture require?

Nuisance does not mean imaginary. Some reported sites are noise or comparison artifacts; others are real physical variations that are not relevant to the current control objective. Their electrical significance depends on defect type, size, location, and the product. A feature dismissed for yield monitoring may still matter to reliability. Review representative candidates before suppressing a class, and retain a check on what the recipe is discarding.

Optical inspection locates; review investigates. High-resolution electron-beam review can revisit selected coordinates to reveal morphology and help classify bridges, breaks, residues, or particles. Inspection and review should be evaluated together: a highly sensitive scan is less useful if nuisance volume consumes the review capacity. The reviewed sample must also be chosen deliberately; reviewing only the strongest signals can hide poor capture or misclassification near the threshold.

Compare results only with the recipe and coverage attached. Record the inspected layer, optical mode, care areas, excluded regions, sampled dies, thresholds, and reference strategy. A lower defect count after narrowing the inspected area is not evidence of a cleaner process. Monitor repeatability and capture on known defects, then examine spatial signatures and trends against process history. Brightfield inspection becomes useful process control when the reported sites can support a decision, not merely when the wafer map contains many dots.

Sources: Applied Materials, Enlight 2 and HawkEye Optical Inspection; KLA, Broadband Plasma Inspection history and Process Watch automotive defect monitoring; KLA patent US10151706B1, Inspection for Specimens with Extensive Die-to-Die Process Variation; Nikon, reflected-light microscopy components. Illustrations explain principles rather than vendor specifications.