In-Line Metrology
In-line metrology measures selected wafer properties at defined points in the manufacturing flow so that process decisions can be made before more value is added. The result might be a film thickness, critical dimension (CD), overlay displacement, or electrical property. Its value comes from connecting a trustworthy measurement to the correct wafer, process history, and control action—not simply from collecting more numbers.
In-line is a place in the workflow, not a promise of instantaneous measurement. A wafer can leave a process chamber, visit a standalone metrology tool, and return to production. Some measurements are integrated into a track or other equipment. In-situ measurements, by contrast, observe the process inside the processing environment. A result delivered after the next operation may still help future wafers, but it cannot correct an operation already completed on the measured wafer.
A sampling plan decides what the fab can see. Measuring every feature on every wafer at every step is generally impractical. Engineers select lots, wafers within a lot, and sites within each wafer. Those choices should reflect the failure modes: center-to-edge variation, localized signatures, chamber differences, and changes over time. A stable average from a sparse map does not establish that all unmeasured regions are good.
The measurement method must fit the question. Optical methods can infer thickness or patterned dimensions from light interacting with a structure; CD-SEM measures selected dimensions from electron images. Overlay metrology measures registration between layers. Electrical measurements address quantities such as resistance. Each method has its own targets, sensitivity, assumptions, and throughput. A target in a scribe lane is useful only to the extent that it represents the behavior the control system needs to manage. Inspection locates or flags potential defects; it complements dimensional metrology rather than being interchangeable with it.
Repeatability is necessary, but it is not accuracy. A tool can return nearly identical readings while remaining offset from a suitable reference. Calibration, reference stability, measurement uncertainty, and tool-to-tool matching all matter. If wafers move between two unmatched tools, an apparent process shift may actually be a measurement shift. Matching must consider the relevant products and measurement conditions; agreement on one reference does not prove agreement on every structure.
Feed-forward changes a later operation on the measured material. For example, a validated controller may use incoming film thickness to adjust a downstream process within an approved operating range. The measurement must arrive before the decision, and the controller must know which wafers it describes. Feed-forward is not permission to compensate for any upstream defect: some conditions require a hold, rework, or rejection instead.
Feedback changes future executions of a process. A post-process result can update a recipe for later wafers or lots. Run-to-run control is one implementation. The delay matters: several lots may already have passed through the tool before an excursion becomes visible. Feedback should not chase measurement noise or mix incompatible chamber and product histories. The diagram separates these two directions of control; real fabs may combine them in a coordinated strategy.
A number needs context before it becomes a control signal. Useful records associate the result with wafer and lot identifiers, site coordinates, units, timestamp, measurement recipe, tool, and processing history. Model-fit quality or other validity flags can be as important as the reported value. Missing, stale, or invalid data must not silently become a normal recipe correction. A controlled workflow defines the fallback: additional measurement, an engineering review, a hold, or an explicitly qualified default.
Monitoring and automatic correction are different decisions. Statistical process control detects changes relative to established process behavior; specification limits describe what is acceptable for the intended requirement. Neither a single noisy point nor an average inside specification is enough, by itself, to justify a recipe change. Engineers investigate whether a signal comes from the process, the measurement system, sampling, or a changed product mix before applying the appropriate response.
The practical objective is timely, actionable information. More sites can improve spatial coverage but consume tool time and delay decisions. Fewer measurements can increase uncertainty or miss excursions. A sound plan balances detection capability, measurement uncertainty, cycle time, and the consequence of a missed problem. In-line metrology succeeds when a reliable result reaches the right decision while there is still an opportunity to act.
Sources: NIST Engineering Statistics Handbook—Accuracy and Bias, and Calibration; ASML—Measuring Accuracy and YieldStar 375F; AMD—WO2003044851A2, feed-forward semiconductor process control.