YieldOpsAcademy
Zone 00 · Clean Room

The Wafer's Journey

Follow Wafer #8847 as a quartz focus ring erodes at 0.3nm per RF hour. Every individual sensor reading is within spec. The process is failing anyway. Master CUSUM through the failure mode that threshold-based monitoring cannot see.

🔍
Today's Subject
Drift Detection
CUSUM for slow manufacturing drift in semiconductor processes

The Memory Problem

Wafer #8847 is the 1,847th wafer processed in Chamber 12 since the last preventive maintenance cycle. The quartz focus ring has been eroding at 0.3nm per RF hour for weeks. Each individual wafer is fine. Each individual sensor reading is within its control limits. A Shewhart chart shows green on every parameter. But something is accumulating. The etch depth on edge dies has shifted by 1.4nm over the last 200 wafers. By wafer 8847, it will be 2.1nm. The device spec is 2.5nm. There are 400 wafers left in the queue.

01
The Physics of Slow Drift
Quartz focus rings shape the plasma sheath near the wafer edge. As the ring erodes, the sheath geometry changes subtly. The tool compensates: matching network capacitors adjust, backside helium pressure creeps up. These compensations are small enough to stay within individual SPC limits but consistent enough to cause monotonic etch depth drift.
02
Why Shewhart Charts Miss It
A Shewhart control chart is memoryless. Each sample is evaluated independently against a fixed threshold. A drift of 0.08 standard deviations per wafer never trips a 3-sigma limit until the cumulative shift reaches 3 sigma, which takes 37 wafers. By then you have already scrapped the lot.
03
CUSUM Accumulates the Evidence
CUSUM maintains a running sum of deviations from the target. Each small-but-consistent deviation from wafer to wafer contributes to the statistic. Five wafers into the drift, the accumulated evidence crosses the control limit and fires the alarm. The drift has been caught 32 wafers before a Shewhart chart would notice.
04
The Response: PM, Not Retraining
The CUSUM alarm routes to the process engineering team, not to the data science team. The correct response is to schedule a preventive maintenance cycle to replace the focus ring. Retraining the yield model on drifting data would hide a hardware failure behind a software fix.

CUSUM does not care what caused the drift. It cares that the process is consistently trending away from its target. The physics investigation is the engineer's job. CUSUM's job is to tell them when to start.

Full Access Required

Continue the journey

Zones 01 through 04 cover the problem scenario, algorithm analysis, alternative comparisons, interview gauntlet, and production checklist for this journey.

All six journeys are included with full access.

Unlock full access · $149See the learning path