precision

Precision, in metrology, measures how tightly repeated measurements of the same feature agree with each other — a question entirely separate from whether those measurements are close to the true value, which is accuracy's job instead. ISO 5725, the international standard that formalizes this distinction, defines precision purely in terms of scatter: a critical-dimension SEM measuring the same transistor gate ten times and getting ten nearly identical readings is precise regardless of whether the tool's calibration happens to be off by a fixed amount, because precision only asks how consistent the readings are with each other, not with reality. At advanced nodes, where overlay tolerances sit in the low single-digit nanometers, this distinction stops being academic: a tool can be imprecise enough that its own measurement noise exceeds the entire error budget a process engineer is trying to control.

Precision and accuracy fail independently, and a tool can have either problem without the other. A measurement system can be precise but inaccurate — consistently reporting a film thickness 0.3nm thinner than the true value every single time, tightly clustered but systematically off, which a calibration correction fixes without touching the underlying consistency; or it can be accurate but imprecise — centered correctly on the true value on average, but scattering widely enough on any individual reading that no single measurement can be trusted. Statistical process control depends specifically on precision rather than accuracy, because SPC charts are built to detect real process shifts, and a control chart built on noisy, imprecise measurements can't tell a genuine shift in the process from ordinary measurement scatter.

Repeatability and reproducibility are the two halves of precision, and conflating them hides where a measurement problem is actually coming from. Repeatability is short-term precision — the same operator, measuring the same feature, on the same tool, in rapid succession; reproducibility is precision across changed conditions — different operators, different tools, different days, or in the strictest sense, different labs measuring the same sample entirely. A fab can have excellent repeatability (one tool, one operator, tightly consistent) while suffering poor reproducibility (that same feature measured differently tool-to-tool), which is exactly the failure mode Gauge R&R — repeatability and reproducibility, measured together as a formal study — exists to isolate, following the methodology laid out in the AIAG's widely adopted Measurement Systems Analysis manual.

The standard rule of thumb treats precision as a tax on the tolerance budget, and the tax rate matters enormously at advanced nodes. A common guideline, following AIAG MSA practice, is that a measurement system's precision should consume no more than 10 to 30 percent of the total tolerance it's being used to control — a CD-SEM tool with a specified precision under 0.5nm (3σ) measuring a feature with a ±2 to 5nm tolerance window is comfortably inside that budget, while the same 0.5nm of noise applied to an overlay tolerance that has shrunk to ±2nm at the leading edge eats a much larger share of the available margin, which is exactly why metrology tool precision requirements tighten in lockstep with every process node shrink rather than staying fixed.

Averaging multiple measurements is the standard, provable way to buy back precision a single reading doesn't have. Because random measurement noise shrinks in proportion to the square root of the number of repeated measurements, averaging 9 readings of the same feature reduces the noise by a factor of 3 compared to a single reading, and this square-root relationship is why fabs routinely specify multi-site or multi-measurement averaging on critical steps rather than trusting any one reading — it's a direct, quantifiable trade of measurement time for measurement confidence, not a rule of thumb.

Precision's real function in a semiconductor fab is as a gatekeeper: no amount of sophisticated process control, SPC charting, or yield analysis can be more trustworthy than the measurements feeding it, and a metrology tool whose precision isn't a small, known fraction of the tolerance it's meant to police isn't measuring the process — it's measuring its own noise and calling it data.