Fab-Wide Control

Fab-wide control connects equipment behavior, wafer measurements, and manufacturing history so that a factory can detect variation, contain excursions, and coordinate process corrections across tools and steps. A chamber controller sees its own operation. A fab-wide system adds the context: which wafers ran there, what happened upstream, where those wafers went next, and whether the final results changed.

The foundation is wafer genealogy, not a dashboard. Genealogy is the traceable history of a wafer or lot through processing, measurement, splits, merges, and rework. A thickness result needs a wafer identity, site, unit, timestamp, and measurement recipe; the associated process record needs the actual tool, chamber, recipe revision, and run. Joining records only by approximate time or lot name can connect the wrong result to the wrong operation.

Different systems answer different questions. The manufacturing execution system (MES) manages material state and permitted routes. Fault detection and classification (FDC) examines equipment signals for abnormal behavior. Statistical process control (SPC) monitors changes in measured process behavior. Advanced process control (APC), including run-to-run control, calculates qualified recipe adjustments. Yield analysis relates process history to defects and electrical outcomes. Fab-wide coordination connects these functions; it does not require every raw trace to reside in one central MES database.

Cross-step control must respect the physical process. Incoming film thickness can inform a downstream polish-time calculation, while post-process measurements can adjust future runs. Such corrections require a validated relationship between recipe inputs and outputs, an allowed adjustment range, and a defined owner. Two independent controllers acting on the same error can over-correct. A feed-forward model should not conceal an upstream fault that requires a hold rather than compensation.

Fleet averages can hide the problem. Suppose two chambers process equal numbers of otherwise comparable wafers. One runs above target and the other equally below it: their combined mean is exactly on target. Grouping by chamber exposes the offsets. Product, layer, recipe, maintenance state, and measurement tool can also matter. Compare equivalent populations before declaring tools matched, and check metrology bias before attributing a difference to the process.

Excursion containment is a genealogy problem with a clock attached. An alarm may arrive after several more lots have used the suspect chamber. The response needs to identify potentially exposed material, including wafers already downstream, and prevent further exposure under the approved procedure. A last verified good result helps establish a review boundary, but it does not prove the exact instant when the fault started. Investigators refine the scope using traces, measurement history, maintenance events, and additional checks.

Correlation narrows an investigation; it does not establish causation. Poor electrical results might share a chamber, but they might also share an incoming material batch or a difficult product. Start with comparable populations and preserve the time order of events. Confirm a suspected mechanism through engineering analysis and appropriate controlled checks before changing a recipe across the fleet. A factory-wide rollout magnifies both a correct diagnosis and a mistaken one.

Virtual metrology adds predictions, not direct measurements. Models can estimate wafer properties from equipment signals and other context, helping prioritize physical measurements and support qualified control. Those predictions need validation against measured results and monitoring for drift. Maintenance, new products, or changed sensor behavior can invalidate a learned relationship. Uncertain or out-of-domain predictions should follow a defined fallback rather than silently replacing trusted measurements.

Coordination also needs limits. Recipe revisions, model versions, correction bounds, material holds, and release authority should be traceable. Missing data and delayed results need explicit handling. Local equipment safety interlocks remain essential; a fab-level analytics service is not their replacement. Successful fab-wide control means the right material is protected and the right process is corrected, with evidence that variation actually decreased, not merely that every tool appears on a common screen.

Sources: Applied Materials, E3 Equipment and Process Control; Onto Innovation, Discover Run-to-Run; NIST/IRDS, Virtual Metrology White Paper; NIST Engineering Statistics Handbook, semiconductor process characterization.