Advanced Process Control

Advanced process control (APC) uses measurements, process models, and qualified adjustments to keep semiconductor manufacturing outputs near their targets. Instead of using the same recipe indefinitely, a controller accounts for predictable changes such as chamber drift or incoming film thickness. The objective is reduced variation within an approved process window, not unrestricted automatic optimization.

The model determines the correction direction. An output error alone does not tell the controller which knob to turn. A useful local approximation is y = a·u + b, where y is the output, u is a recipe input, a is the measured process sensitivity, and b is an offset. Ignoring other changes, the predicted correction is Δu = (target − measured output)/a. The units and sign of a matter. A positive output error calls for decreasing u when a is positive, but increasing u when a is negative. Lithography dose-to-CD behavior must be established for the relevant pattern, resist process, and measurement stage; there is no universal dose correction for every high-CD result.

Feedback and feed-forward use different information. Feedback uses an observed process output to improve future runs. Feed-forward uses information about the incoming wafer to choose a later operation on that wafer. In a simplified CMP example, remaining thickness equals incoming thickness minus removal rate times polish duration. If the removal rate is stable, a thicker incoming film requires more polishing to reach the same remaining thickness. Real CMP control must also account for material transitions, endpoint behavior, uniformity, and approved limits; the simple equation is an explanation, not a production recipe.

Run-to-run control does not require instantaneous metrology. Measurements may arrive between wafers or lots, sometimes after other material has already been processed. A controller must associate each result with the recipe that produced it and account for that delay. Applying an old error as if it came from the latest recipe can produce the wrong correction. Separate control histories, often called threads, keep qualified product, layer, tool, and chamber combinations from contaminating each other's estimates.

Filtering trades speed for noise rejection. An exponentially weighted moving average (EWMA) can update an estimated process offset: new estimate = λ·new observation + (1 − λ)·previous estimate, with 0 < λ ≤ 1. A larger λ responds more strongly to recent data; a smaller one smooths more but follows a sustained shift slowly. This is an estimation rule, not a universal recipe-update equation. Model inversion, delay handling, and adjustment limits still determine the commanded recipe.

More than one knob may affect more than one result. Changing an etch input can affect depth, CD, and profile together. Multi-input, multi-output control uses a model of these coupled responses to choose a feasible adjustment, rather than optimizing one measurement while degrading another. A model trained in one operating region may not remain valid after a recipe change or maintenance event. Model versions and qualification boundaries therefore belong in the control record.

An equipment fault is not ordinary drift. Fault detection and classification (FDC) can flag abnormal runs so their results do not automatically update the controller as if they represented normal process behavior. Missing measurements, failed metrology quality checks, stale data, or excessive requested corrections need an explicit response: freeze adaptation, use an approved fallback, request measurement, or hold material. Controller limits do not replace equipment safety interlocks.

Qualification tests the closed loop, not just prediction accuracy. Engineers check model sensitivity, measurement uncertainty, noise response, delay, abrupt shifts, and actuator limits before enabling automatic correction. Historical replay and advisory operation can precede a limited production release. After deployment, both output variation and recipe behavior should be monitored: a mean near target can hide oscillation, excessive corrections, or poor within-wafer uniformity. APC is useful when measured performance improves without moving the process outside its qualified window.

Sources: Onto Innovation, Discover Run-to-Run and Smart Manufacturing Software; NIST Engineering Statistics Handbook, EWMA Control Chart; Stability Analysis of Semiconductor Manufacturing Process with EWMA Run-to-Run Controllers (arXiv:1510.08946).