Raw data
Everything about one wafer
Not only the corrected result, but the raw waveforms from both probes and the position data are kept together, so the same wafer can be recomputed later a different way.
AI Roadmap
The tool improves itself from the data it already writes
ZAFS already records the raw scan waveform, alignment diagnostics and motion history for every single wafer. Until now that record was something an engineer opened after a problem. From here it becomes the material for reading a problem before it happens.
Let physics carry what physics explains.
Give the data only what is left.
What we already have
The starting point is data, not an algorithm
The most common way to fail at putting AI into equipment is to choose the model first, with no data to train it on. ZILAB came at it in the opposite order.
Diagnostics
Alignment snapshots
Whether the alignment succeeded or failed, the edge signal and the intermediate estimates from that moment are stored with it. Failures can be reproduced exactly afterwards to narrow down the cause.
Physics + sim
Physical models and a simulator
The physical models for gravity sag and stage reference plane already stand, and a simulation layer runs the whole process without hardware. There is somewhere to verify a learned result before it reaches a real tool.
Four stages
Roadmap
Not easiest first — verifiable first. Each stage assumes the data and the confidence the previous one produced.
-
STAGE 1
Announce the stop — anomaly prediction
Alignment failures and transfer stops do not arrive without warning. Leading indicators begin to drift tens of minutes before the failure. ZILAB has already confirmed, by analysing failures after the fact, that indicators which move ahead of the failure exist. Learning them lets the data set the threshold instead of a person.
What it buys: an unplanned stop becomes a planned one. Letting the customer choose when the line goes down matters more than the uptime figure itself.
-
STAGE 2
Learning on top of physics — hybrid correction
Gravity sag and the stage reference plane are explained by physical equations. But a real tool leaves a residual those equations do not cover — assembly tolerance, temperature, long-term drift. Only that residual is given to the learned model.
What it buys: it does not collapse when training data is scarce, and outside the training range the physical term still holds the value in place. Tool-to-tool variation and long-term drift both shrink.
-
STAGE 3
The tool finds its own parameters
Scan velocity and acceleration, alignment window width, sampling interval, wait times — dozens of parameters govern performance, and today a person sets them from experience. With measured TACT and scatter as the objective, they are searched automatically on the simulator and then confirmed on a real tool.
What it buys: throughput and repeatability rise together without changing the hardware. Each tool finds its own optimum.
-
STAGE 4
Change the development process itself
AI assistance goes into the chain from requirement to code, code to regression test, test to documentation. Hundreds of automated regression runs on the simulator mean the code that reaches a real tool has already passed verification.
What it buys: shorter release cycles and fewer regressions. Less time between a customer requirement and the tool doing it.
Expected quality gain
Where the product actually gets better
Below is the point each stage changes. Rather than promising numbers, it states what becomes different and how.
| Area | Today | With AI |
|---|---|---|
| Unplanned stops | Respond after an alignment or transfer failure | Warn early from leading indicators; convert to a planned stop |
| Tool-to-tool variation | Held inside specification by master calibration and tool matching | Residual learning narrows it further and tracks long-term drift automatically |
| Repeatability | Standard deviation 0.002 – 0.003 µm under ideal conditions | Condition-aware correction reduces scatter under real production conditions |
| Throughput | 55 sec/wafer on fixed, human-set parameters | Per-tool automatic search recovers margin from the same hardware |
| Anomaly detection | Reject wafers outside specification | Report wafers inside specification that look unlike the norm, as a process signal |
| Software quality | Checked by a person on a real tool | Caught before release by automated regression on the simulator |
Principles
Three rules for putting AI into a metrology tool
01
It does not replace physics
Hand the explained part over to learning as well and the value collapses wherever there is no data. Learning is used only on the residual physics does not explain.
02
No decision without a reason
A metrology decision is a decision to discard a customer wafer. A result whose reasoning cannot be traced back through the data is not emitted.
03
Learn on the tool, verify on the tool
Both training and verification use data a real tool produced. Run it on the simulator first, confirm it on a real tool, and only then make it the default.