Wafer excursions investigated in hours. Every finding linked to its lot.
A global semiconductor manufacturer connected wafer maps, lot genealogy, tool history, metrology and past incidents, so an excursion gets a ranked, evidenced set of hypotheses in one run and the engineer decides what happens to the lots.
- Customer
- A global semiconductor manufacturer
- Scope
- Private cloud · two fabs
- Published
What was in the way
A yield excursion is a race. Every hour spent finding the cause is another set of lots through the same tools. The evidence existed, but it lived in five places: probe results and wafer maps, the lot genealogy in the MES, equipment logs and chamber state, inline metrology, and an archive of past incident reports written in free text.
Pulling those together fell to whichever engineer was on shift. The first credible hypothesis often took most of a day, and the reasoning behind it lived in a slide deck that was hard to audit later.
The fab had tried dashboards that correlated everything with everything. They produced plenty of candidates and very little confidence, because nobody could see why a tool had been flagged.
How Agent Hive was set up
- 01
Connect the evidence read-only
Probe data, wafer maps, lot genealogy, equipment history, metrology and the incident archive were connected read-only under the fab’s existing access rules. Nothing leaves the private cloud. - 02
Describe the fab once
Routes, steps, tools, chambers and recipes were described in a reviewed knowledge graph, so an agent knows which lots shared which chamber at which step without re-deriving it every run. - 03
One investigation run
A primary agent classifies the wafer signature, gathers the affected lots and delegates tool, metrology and history analysis to specialists. Their findings come back with the records attached. - 04
Rank, do not decide
Hypotheses are ranked with supporting and contradicting evidence. Recommendations to hold, inspect or release go to the owning engineer, who approves, edits or rejects them. - 05
Grade against past excursions
Closed excursions with known causes became the evaluation set. A change to the workflow ships only if it still finds those causes.
What changed
- Time to first hypothesis
- 3.5 h
- median, was about a day
- Findings with sources
- 100%
- lot, tool or record linked
- Closed excursions in the eval set
- 40
- graded before each change
- Lots moved by an agent
- 0
- engineer approves every hold
Start from the signature
An investigation opens when a yield monitor fires or an engineer asks a question: “Why did bin 7 jump on these lots?” The first step is always the wafer map. The agent classifies the spatial signature (edge ring, centre cluster, scratch, repeating reticle pattern) and uses it to narrow where to look. An edge ring points towards deposition and etch uniformity. A scratch points towards handling.
Wafer map with defect signature. Wafer map of W07 in lot L-2417: 352 dies, 90 fail, yield 74.4%. 78 failing dies form a edge ring signature; the classifier reads edge ring at 0.86 and links the wafer to investigation INV-20418 and a hold under SOP-QA-001.
Every hypothesis carries its evidence
From the signature, the primary agent pulls the affected lots and their genealogy, then hands sub-questions to specialists: one compares tool history and chamber state for the steps those lots shared, another checks inline metrology against limits, a third searches the incident archive for past excursions with the same signature. Each returns findings with the records behind them.
RCA investigation trace. Investigation INV-20418, anchored on: Etch depth above the control limit (206.7 nm · UCL 206 nm; RIE-02 · L-2417 on hold). The RCA Investigator agent worked 7 steps: 1. Prior knowledge: FailureMode: ICP etch · focus ring wear; 2. Query lot events: LotEvent: LEV-88102…88107 · L-2417; 3. Read state history: Machine: RIE-02 · state history, last PM 418 h ago; 4. Search SOP: Document: SOP-ETCH-004 p.3 · PM every 400 h; 5. View wafer map: Image: L-2417 W07 wafer map · edge ring; 6. Twin check: Twin: etch rate model · live above prediction by 3.1 %; 7. Compare tools: Run: E-02_P118 · RIE-01, ETCH-NB-04 v3, in limits. Hypotheses: H1 Focus ring worn on RIE-02, confirmed; H2 Recipe ETCH-NB-04 v3 change, refuted by Run: RIE-01 · same recipe, in limits; H3 Resist thin at the wafer edge, refuted by Measurement: pre-etch map uniform; H4 Metrology drift on MET-03, refuted by Measurement: reference wafer in limits. Root cause proposed by the agent: Worn focus ring on RIE-02, confirmed by a process engineer. Solution waiting as a pending proposal: PM interval 400 → 320 h (on SOP-ETCH-004 · approval requested). L-2417 held · retest edge dies, disposition off-flow under SOP-QA-001. Verified by a person: Next 3 lots in limits (L-2418, L-2419, L-2421).
The output is a ranked list of hypotheses, each with the lots, tools, measurements and earlier incidents that support or weaken it. Dead ends are kept. When an engineer asks why a chamber was ruled out, the run shows the comparison that ruled it out.
The engineer decides
Agents never move a lot. A recommendation to hold, inspect or release is routed to the owning yield engineer with the evidence attached. The engineer approves, edits or rejects it, and that decision is stored with the run. Rejections with a reason are the most valuable signal the workflow gets: they become cases the next version is graded against.
The useful part is that I can see why it thinks a chamber is involved, and what it checked before ruling the others out.