process control
**Process control in semiconductor manufacturing is the system that keeps thousands of interacting fabrication steps centered inside narrow specifications.** Statistical process control (SPC) detects unusual variation, advanced process control (APC) adjusts recipes, fault detection and classification (FDC) monitors equipment traces, and excursion management contains risk. Together they convert measurements and tool signals into decisions before small drift becomes lost yield across many wafers.
**Variation comes from tools, materials, environments, designs, and measurement itself.** Chamber walls season, consumables wear, sensors drift, incoming films vary, and product patterns load processes differently. Common-cause variation is the stable background of a capable process; special-cause variation signals an identifiable change. Treating every fluctuation as a fault creates needless adjustments, while ignoring a real shift sends defects downstream where they become more expensive.
| Method | Primary question | Inputs and cadence | Typical action |
|---|---|---|---|
| SPC | Is output statistically stable? | Sampled dimensions, film and defect data | Hold, investigate, change control limits |
| Run-to-run APC | What recipe correction should the next lot receive? | Metrology plus process context after each run | Adjust dose, time, focus, pressure, or zones |
| FDC | Did the equipment behave abnormally during this wafer? | High-rate sensor and subsystem traces | Alarm, classify, stop chamber, request maintenance |
| Virtual metrology | What is likely output without direct measurement? | Tool traces, context, trained model | Predict quality and increase control frequency |
| Predictive maintenance | When will a component degrade? | Vibration, RF, vacuum, temperature, history | Schedule service before failure or drift |
**Control charts separate signal from expected noise.** A center line represents the estimated process mean, and control limits reflect natural variation rather than drawing specification limits. Western Electric-style rules flag points beyond limits, persistent runs, trends, or unusual clustering. Charts may track individuals, moving ranges, means, ranges, proportions, or defect counts. Limits must be recomputed only through governed baselines; automatically widening them after an excursion conceals instability.
```svg
```
**Capability compares process spread with product specifications.** For a centered normal process, \(C_p=(USL-LSL)/(6\sigma)\), while \(C_{pk}\) also penalizes an off-center mean. Fabs often seek Cpk values above 1.33 and higher for critical mature characteristics, but non-normal distributions, autocorrelation, sampling bias, and measurement error can make a simple number misleading. Stability must be established before capability is interpreted.
**Run-to-run control compensates predictable drift.** An exponentially weighted moving average controller estimates process state from recent metrology and changes the next recipe. Lithography adjusts dose, focus, alignment, or wafer-level corrections; deposition changes time or precursor conditions; CMP changes polishing time or carrier zones. Feed-forward uses incoming measurements, while feedback uses output. Controller gain must avoid oscillation when metrology arrives late.
**FDC evaluates the complete equipment trace.** Pressure, RF power, gas flow, temperature, valve position, endpoint, robot timing, and subsystem status are aligned to recipe phases. Limits and multivariate models detect deviations that final metrology might miss. Classification links trace signatures to leaks, arcing, worn consumables, sensor problems, or handling events. Excessive alarms teach operators to ignore the system, so alarm quality and actionability are controlled metrics.
**Virtual metrology predicts results between physical measurements.** Models combine trace features, chamber history, product context, and recent actual metrology. They can increase effective sampling and route suspicious wafers to confirmation, but prediction uncertainty and domain drift must be reported. A model trained before maintenance or on one product may fail afterward. Periodic ground-truth sampling remains essential.
**Measurement systems need their own control.** Gauge repeatability and reproducibility, matching among tools, calibration, sampling location, and recipe version affect observed variation. If measurement noise consumes a large share of tolerance, a controller may chase the gauge rather than the process. Reference wafers, golden tools, correlation studies, and automated health monitors separate metrology drift from process drift.
**Excursion management limits blast radius.** When a rule triggers, systems identify affected chamber, time window, lots, wafers, layers, and downstream dependencies. Material is held while engineers review traces and measurements. Decisions include release, additional inspection, rework, downgrade, or scrap. Fast genealogy queries matter because a faulty shared tool can touch many products before a delayed test reveals the issue.
**Chamber matching enables flexible factory scheduling.** Nominally identical chambers have distinct offsets and histories. Golden-chamber comparisons, matching wafers, and per-chamber APC corrections reduce those differences. Qualification after preventive maintenance demonstrates that a chamber has returned to baseline. Without matching, dispatch systems lose flexibility and products accumulate queues at favored tools.
**AI and machine learning add pattern recognition but not automatic trust.** Models can predict maintenance, classify defects, detect multivariate anomalies, optimize sampling, and connect inline signals to final yield. Rare failures, changing recipes, and incomplete labels are serious constraints. Engineers need explainable features, uncertainty, versioning, shadow evaluation, rollback, and human escalation. A false negative may expose thousands of wafers; a false positive may stop a bottleneck tool.
**Data architecture is part of control performance.** Sensor streams need synchronized timestamps, units, recipe phases, equipment identity, product context, and lineage. Metrology and yield results may arrive days or weeks later. Feature stores and historians must preserve versions so a past decision can be reconstructed. Cybersecurity prevents unauthorized recipe changes while allowing low-latency approved controllers to act.
**Governance distinguishes a validated control from an experiment.** Owners define targets, limits, model scope, fallback, approval, access, and response plans. Changes are tested on engineering material or in shadow mode, then released with monitoring. Operators receive alarms that state what happened, which material is affected, and what action to take. Audit trails record automated corrections.
**Effective process control reduces both mean loss and uncertainty.** It raises yield, shortens learning cycles, prevents escapes, improves equipment utilization, and provides evidence for root cause. The strongest fab does not eliminate variation; it measures the right variables, recognizes meaningful change early, applies bounded corrections, and learns from every excursion without destabilizing a healthy process.
**Sampling strategy allocates scarce measurement time to risk.** Fixed sampling is simple but may miss a short excursion or overmeasure stable layers. Dynamic sampling raises coverage after maintenance, recipe change, abnormal trace, new product introduction, or weak capability, then relaxes it after evidence of stability. Wafer and site selection must represent center-to-edge and within-lot behavior. Skipping measurement based on a virtual model requires explicit uncertainty thresholds and periodic forced samples so the model cannot silently drift away from reality.
**Final electrical test closes a long feedback loop.** Inline dimensions and defects are leading indicators, while wafer-sort signatures show which process variation actually affected circuits. Engineers link spatial fail maps and parametric distributions to chambers, routes, masks, and upstream measurements. Because that feedback arrives late, causal analysis must account for product mix and confounders. Lessons are converted into tighter monitors or new control variables so the same yield mechanism is caught earlier on future lots.