AI yield copilot for semiconductor fabs

Catch the excursion
before the next lot.

Sentry reads every wafer map as it leaves the tool, notices when one tool keeps producing the same defect, and tells the engineer the likely cause, the lots at risk and what it costs.

Open the live demo See how it works
The Sentry operations view: tool status, a live inspection feed and an Edge-Ring alert on ETCH-04 with its likely cause and exposure
0.912macro-F1 across 9 defect patterns
95%of defective wafers caught
~10 msper wafer on a CPU
811,457real wafer maps in WM-811K

Model scores are on the held-out validation lots. The test set is kept back for one final evaluation.

Built on WM-811K wafer maps AMD Token Factory PyTorch FastAPI

The problem

The defect is on the map.
It is found too late.

After test, every wafer gets a map of which dies passed and which failed. The shape of the failures points at the process step that caused it: a ring at the edge, a scratch, a cluster in the centre.

Those maps are checked by hand, in batches, or not at all. By the time someone notices that one tool keeps producing the same shape, it has already run more lots.

Sources: WM-811K (Wu, Jang and Chen, IEEE Transactions on Semiconductor Manufacturing, 2015). Wafer value is the demo's stated assumption.

  • 85%

    Most maps are clean

    Of the 172,950 labelled maps, 85% show no defect pattern. The few that matter are buried among them.

  • 9

    Patterns, nine different causes

    Edge rings, scratches, centre clusters and donuts each point at a different tool or process problem.

  • $500k

    At risk per lot

    25 wafers at an assumed $20,000 each. Every lot a drifting tool runs before someone acts adds to it.

The copilot

From wafer map to decision in one view.

Every wafer is classified as it leaves the tool. When a pattern repeats on one tool, the copilot raises an alert, explains it and puts the decision in front of the engineer.

  • Classifies every wafer

    A calibrated CNN ensemble names the defect pattern in about 10 ms. When it is unsure, the wafer goes to a person instead of guessing.

  • Spots the excursion

    An alert fires when one tool produces the same pattern three times within 24 hours.

  • Explains the likely cause

    An LLM copilot on AMD Token Factory pulls the lot history, likely root cause and cost, and cites where each answer came from.

  • The engineer decides

    Hold the lots, open a ticket or dismiss. Every decision is logged with the time and who made it.

ExcursionFrom the demo line

Edge-Ring on ETCH-04

3 wafers in 24 h · fired 00:06 in LOT-010

Edge-Ring wafer map
Edge-Ring
Edge-Ring wafer map
Edge-Ring
Edge-Ring wafer map
Edge-Ring
Lots run since alert4
Exposure since alert$1.62M

Likely cause · triage, not diagnosis

Edge temperature or plasma non-uniformity. Check focus and edge ring wear and edge etch-rate uniformity.

Hold lotsOpen ticketDismiss

How it works

From the test floor to a held lot

Three steps, each one you can inspect in the demo.

  1. Center
    Donut
    Edge-Ring
    Edge-Loc
    Loc
    Random
    Near-full
    No pattern
    01

    Inspect

    Each wafer map is read the moment it leaves the tool. The model sorts it into one of eight defect patterns or no pattern, and states how confident it is.

  2. Scratch wafer map
    Wafer map
    The same wafer with the model's attention highlighted on the scratch
    Model attention
    02

    Explain

    Every call can be opened: the map, the area that drove the decision, and the score for all nine patterns. Uncertain wafers are flagged for a person.

  3. Alert raisedEdge-Ring ×3 on ETCH-04 · 00:06
    Copilot briefedEdge ring wear likely · 4 lots exposed
    Lots heldby the shift engineer · 00:21
    03

    Act

    The engineer holds the lots, opens a maintenance ticket or dismisses the alert. The copilot recommends; a person decides.

Measured honestly

Scored on lots it has never seen.

A model can look perfect when wafers from the same lot end up in both training and testing. We split by lot, so every score below comes from lots the model never trained on.

  • Grouped by lot. No lot appears in more than one split.
  • Calibrated. When it says 90%, it is right about 90% of the time.
  • Knows when to ask. Under 1% of wafers go to a person for review.
  • Test set kept back for one final evaluation.
Read the model card
F1 score by defect patternValidation lots · v0-ens3

F1 combines how many of each pattern the model finds with how many of its calls are right. 1.0 is perfect.

Who it's for

Built for the people who own yield

Yield and process engineers

Find the tool behind the pattern

  • Every wafer classified, not a weekly sample
  • Likely cause with sources for each alert
  • Ask follow-up questions in plain English

Fab and shift managers

See the cost before it grows

  • Exposure in dollars as lots keep running
  • One view of every tool on the line
  • A log of who decided what, and when

See it catch an excursion.

The demo replays 500 real wafers through four etch tools. Press play and watch ETCH-04 drift.

Open the live demo View the code