On-Device Overlay

On-device overlay measures registration directly on real, functional device structures rather than on dedicated overlay targets sitting in the scribe line — because at advanced nodes, the gap between what a proxy target reports and what overlay actually looks like at the real device location has grown large enough to consume an entire overlay budget on its own. It's the most accurate, and most difficult, overlay measurement approach available, precisely because it refuses to accept any proxy standing in for the thing that actually matters.

Target-to-device offset — the gap between what a scribe-line target measures and what overlay actually is at the real device location — used to be small enough to ignore, and at advanced nodes it simply isn't anymore. A scribe-line target sits physically separated from the actual device structures it's meant to represent, and any systematic difference in how the lithography and process steps behave between that scribe-line location and the device-interior location shows up as a real discrepancy between measured and actual overlay. At sub-5-nanometer nodes, where total overlay budgets shrink to roughly one to two nanometers, this target-to-device offset can consume a meaningful fraction of the entire available budget before the device itself has even been measured — a gap too large to simply assume away.

Intrafield variation — overlay that genuinely changes across different positions within a single exposure field — is something a single scribe-line target, or even several of them, structurally can't capture. Lens distortion and other field-dependent effects mean overlay isn't necessarily uniform across an entire exposure field; it can vary meaningfully from one part of the field to another. A scribe-line target measures overlay only at its own specific location, telling you nothing directly about how overlay behaves at a different position within the same field where actual devices happen to sit — which is precisely the blind spot on-device overlay measurement exists to close.

Electron-beam-based measurement directly images actual device features at the resolution needed to extract overlay from them, trading speed for the most direct, ground-truth accuracy available. SEM-based overlay measurement can resolve real device structures with the precision needed to extract genuine registration information from them directly, rather than from any proxy target standing in for the device. The real cost is throughput — SEM imaging at this resolution is inherently slower than optical measurement techniques, which is exactly why e-beam on-device overlay tends to be used selectively, for characterization and process-development work, rather than applied to every single wafer in production.

Small in-die targets, placed much closer to actual devices than traditional scribe-line targets, offer a practical middle ground — closing most of the target-to-device gap without needing the throughput cost of full e-beam device imaging. Positioning a compact overlay target within the die area itself, physically near the real devices it's meant to represent, substantially reduces the distance (and the associated process variation across that distance) between what the target measures and what the surrounding devices actually experience. This approach captures much of on-device overlay's accuracy benefit while staying compatible with faster optical measurement techniques, rather than requiring the slower, more resource-intensive e-beam approach for every measurement.

Computational and hybrid approaches extract overlay information directly from design features using pattern matching or machine learning, or combine fast scribe-line measurements with targeted in-die corrections — squeezing more accuracy out of existing measurement infrastructure rather than replacing it outright. Pattern-matching or machine-learning techniques can identify overlay-sensitive features already present in the actual device pattern and extract registration information from them computationally, without needing a dedicated physical target at all. Hybrid schemes take the fast, established scribe-line measurement as a baseline and apply targeted, sparser in-die correction measurements specifically to capture the systematic gap between scribe-line and device-level overlay — a practical compromise that captures much of on-device overlay's real accuracy benefit without abandoning the production throughput scribe-line measurement already delivers.