Semiconductor manufacturing is the most measurement-dense industry on earth. A single wafer passes through several hundred process steps, most of them metrology-monitored, and a mature fab generates control-chart points by the million per day. Fault detection and classification systems watch tool sensor traces at kilohertz. Run-to-run controllers adjust recipes between lots.
Which makes it a strange industry to sell an SPC tool into — and a very good one to learn SPC from, because fabs discovered every failure mode of naive charting decades before anyone else had the data volume to hit them.
The chart most people draw wrong
The defining question in fab SPC is not which chart type — it is what the subgroup is. Sites on a wafer? Wafers in a lot? Lots in a day? Each choice produces a different chart that answers a different question, and the wrong choice makes a real problem invisible.
Wafer (centre, and four sites at radius)
Subgrouped across the wafer, not across time. Not one wafer mean signals — the top chart is flawless from end to end — while from wafer 17 every single range sits above the range chart's upper limit.
The chart that matters here is the bottom one. Within-wafer uniformity has collapsed — edge sites running thin, centre sites thick — so the two errors cancel in the average and the wafer mean stays exactly on target. A chart of wafer means alone would have shown nothing at all.
Illustrative data — generated to behave like this process, not taken from anyone's plant. The limits, the signals and the capability indices above are computed at page load by the same engine that draws a customer's chart.
Here the subgroup is five measurement sites on a single wafer: centre plus four at radius. The X̄ chart plots the mean thickness of each wafer and the range chart plots the spread across the wafer.
From wafer 17 the mean is perfect. Dead on target, comfortably inside its control limits, no signal at all. And the process is falling apart: within-wafer uniformity has collapsed — edge sites running thin, centre sites running thick — so the two errors cancel exactly in the average. The range chart doubles. Every wafer after 17 has a radial signature that will show up at electrical test as a yield ring, and the location chart has nothing to say about it.
This is why range and sigma charts are not decoration. In an industry where "the average is fine" is the most expensive sentence available, the dispersion chart is frequently the one carrying the information.
(This platform does not flag rule violations on the dispersion chart — the limits are drawn and computed, but the rule engine runs on the location chart only. On the chart above the excursion is unmissable by eye; a subtle one would not be. Worth knowing before you rely on it.)
What gets charted
- Litho — critical dimension (CD), overlay/registration, focus and exposure latitude.
- Etch and deposition — etch rate, film thickness, uniformity, selectivity, step coverage.
- Thermal and implant — sheet resistance, junction depth, dose uniformity.
- CMP — removal rate, within-wafer and within-lot thickness range, dishing/erosion.
- Defectivity — particle counts per wafer, defect density per cm², killer defect rates by bin.
- Electrical test / PCM — threshold voltage, drive current, contact resistance, ring-oscillator frequency, plus final yield by bin.
- Equipment — chamber pressure, RF power, gas flows, and the availability metrics defined by SEMI E10.
Why the usual tooling hurts here
Excel. At fab data volumes this is not a discussion.
Enterprise fab SPC (the real thing). Yield management systems, FDC platforms, run-to-run control — these are the correct tools for a production fab and cost accordingly. They integrate with the MES, they handle multivariate models, they trigger tool interlocks. Nothing here competes with that.
JMP, Minitab, Python. Where the interesting analysis actually happens. Excellent for a study, unattended monitoring is not their job.
Where this platform fits
Not in a 300 mm production fab. Where it does fit:
- R&D and pilot lines, where the process is new, the volumes are small, and the enterprise system will not be configured for another year.
- Compound semi, MEMS, photonics, packaging, and specialty houses — smaller fabs where the full yield-management stack has never been affordable and the monitoring is genuinely being done in spreadsheets.
- Equipment and materials suppliers shipping into fabs, who have to demonstrate control of their own process (chamber parts, targets, gases, substrates) with the same rigour their customer applies.
- Incoming and outgoing inspection, where a handful of characteristics need to be watched and shared with a customer.
What you get: subgrouped charts with the subgroup you choose, limits frozen to
a baseline window and versioned, all eight Nelson rules with per-rule severity,
alerting, a read-only wall display, and an API that takes a POST from
whatever is holding your metrology data.
What we do not do
Production fabs are well served by tools we are not trying to be:
- Multivariate SPC. PCA and T² over correlated sensor sets is the standard way to watch a chamber; we chart one characteristic at a time.
- FDC, run-to-run control and interlocks. We observe; we do not act on the tool.
- Rules on the dispersion chart. Its limits are computed and drawn, but the rule engine runs on the location chart, so a subtle range excursion still needs an eye on it.
- Fab-scale volumes. Built for the tens to thousands of points per line per shift that most manufacturers have, rather than millions a day.
- Wafer maps and spatial analytics. The radial signature above shows up as a range excursion; which sites it came from needs a map we do not draw.
Where to start
If your metrology data already lands in a database, the fastest test is one API token and a small script pushing one characteristic — CD, thickness, sheet resistance, whatever hurts most — and letting it run for a week with limits frozen on a stable lot range. The Postman collection is generated from the live routes, so there is no documentation drift to fight.