Beat 01
Process domains have different physics
Deposition drift, anneal response, transfer mechanics and binning economics do not share a loss function. Forcing them into one model buries the signal that matters.
AI Agents
Seven fab-edge agents under one orchestrator
Each agent owns a process domain and is measured on a fab metric. The orchestrator makes sure they do not optimise against each other.
- Domain-scoped
- Orchestrated
- Human-in-the-loop
Agents trained on real display-process signals
Gen-8.6 OLED fabFlexible OLED module lineAutomotive display makermicroLED pilot lineAR/VR microdisplay groupLTPS array fab
Design
Why agents and not one model
Beat 02
Ownership makes accountability possible
Each agent maps to a team that already owns the metric, so an override has an obvious owner and a training label has an obvious source.
Beat 03
The orchestrator holds the trade-offs
A deposition correction that helps uniformity but costs throughput is a fab decision, not a model decision. The orchestrator makes it explicit.
Beat 04
Autonomy is per-agent
An agent can graduate on its own evidence. Mura classification may be trusted long before transfer control is.
# closed loop, one panel
perceive substrate → tft → emitter stack → sub-pixel
plan deposition · anneal · transfer · test/bin
act recipe write-back (SECS-GEM)
sense mura ΔL* · sub-pixel opens · particles
optimise yield · ramp · scrap
log immutable audit entry ✓ signed
retrain engineer correction → model
Orchestration
How a decision travels
The same path every time, whether the outcome is a suppressed nuisance call or a recipe change.
Step 1 of 6 · sense
Agent senses
Domain agent reads its signals from the perception layer.
Step 2 of 6 · propose
Agent proposes
A scoped action with a predicted effect and a confidence.
Step 3 of 6 · arbitrate
Orchestrator arbitrates
Competing proposals are weighed against yield, ramp, scrap and throughput together.
Step 4 of 6 · validate
Twin validates
The proposed action is simulated against panel and line state.
Step 5 of 6 · gate
Autonomy gate applies
Shadow logs it, advisory routes it, graduated autonomy executes low-risk actions.
Step 6 of 6 · record
Audit records
Signed record with evidence, model version and approver.
The seven
Agents and what they own
01
deposit-and-pattern
Holds emitter-stack and TFT layer uniformity across the whole mother glass by closing the loop between chamber telemetry, overlay metrology and recipe write-back.
- Thickness and overlay drift prediction
- Chamber-to-chamber matching
- Recipe write-back over SECS-GEM
02
tft-and-anneal
Stabilises threshold voltage and mobility across the panel so backplane variation never becomes visible non-uniformity at cell test.
- Excimer and thermal anneal control
- Vth / mobility mapping
- Array-test correlation
03
mura-and-defect
Separates real large-area non-uniformity, sub-pixel opens and shorts, particles and stains from the nuisance signals that swamp inspection queues.
- True-versus-false-call classification
- Sub-pixel and particle detection
- Evidence-linked calls
04
microled-transfer
Optimises mass transfer and repair — the yield ceiling on microLED — by learning placement, bonding and repair outcomes emitter by emitter.
- Placement and bonding parameters
- Repair-cycle prioritisation
- Known-good-die routing
05
panel-handling
Coordinates robotic panel and cassette movement so large, fragile substrates move without breakage, particles or queue stalls.
- Breakage and particle risk scoring
- Cassette logistics
- Fail-safe stop to known-good state
06
age-test-and-bin
Runs aging, test and binning decisions with the demand mix in view, so grade calls maximise value instead of defaulting to the safest bin.
- Aging and burn-in analysis
- Grade and bin optimisation
- Demand-aware routing
07
yield-and-ramp
Turns every excursion into a shorter ramp for the next product, compressing the manual tuning cycle that eats new-model margin.
- Excursion root-cause ranking
- New-product ramp playbooks
- Scrap and rework reduction
Accountability
Each agent has a number
If an agent cannot be measured on something the fab already reports, it does not ship.
| Agent | Primary metric | Owning team |
|---|---|---|
| deposit-and-pattern | Uniformity within window / rework | Process integration |
| tft-and-anneal | Vth spread / array-test yield | Array process |
| mura-and-defect | False-call rate / escape rate | Quality and inspection |
| microled-transfer | Transfer yield / repair cycles | Process integration |
| panel-handling | Breakage and particle events | Equipment engineering |
| age-test-and-bin | Recovered value per panel | Manufacturing operations |
| yield-and-ramp | Weeks to target yield | Yield engineering |
Human-in-the-loop
The override is the product
Engineers are not an obstacle to autonomy; they are the training signal that makes it possible. Every override is captured with its reasoning and flows back into the model that made the call.
- Overrides captured with rationale, not just a rejection
- Disagreement between engineer and agent surfaces for review
- Model updates are versioned, rolled out and reversible
- High-impact decisions never graduate out of human approval
# closed loop, one panel
perceive substrate → tft → emitter stack → sub-pixel
plan deposition · anneal · transfer · test/bin
act recipe write-back (SECS-GEM)
sense mura ΔL* · sub-pixel opens · particles
optimise yield · ramp · scrap
log immutable audit entry ✓ signed
retrain engineer correction → model
Under the hood
One orchestrated cycle
This is the loop the orchestrator runs, condensed. In shadow mode every line still executes — only the write-back is withheld.
cycle panel=G6F-118422 mode=advisory
mura-and-defect 3 calls → 1 real, 2 suppressed
deposit-and-pattern drift +0.4nm/h chamber 3
yield-and-ramp excursion rank #1 this shift
orchestrator: correct deposition, hold throughput
twin validate → pass
gate advisory → queued for approval
audit ✓ signed
Where agents connect
Agents read and write through governed interfaces
Deposition, litho and encapsulation
Vacuum depositionInkjet/TFE encapsulationPhotolithographyOverlay metrologyChamber telemetry
Anneal and TFT process
Excimer laser annealThermal annealVth metrologyLTPS/oxide array
Inspection, metrology and test
Mura/AOI systemsOptical uniformityArray testCell testAging
Transfer, repair and handling
Mass-transfer headsLaser repairPanel roboticsCassette logistics
Voices from the line
What the fab floor tells us
“We do not lose panels because nobody is watching. We lose them because the signal that mattered was buried under a thousand nuisance calls.”
“Ramp is the whole game. If a new product takes two quarters of manual tuning, that is two quarters of margin we never get back.”
“Transfer yield is our ceiling on microLED. Every dead emitter is a repair cycle or a scrapped backplane.”
Quotes are illustrative composites drawn from discovery interviews with process-integration, yield and quality engineers. Named references are [PLACEHOLDER] pending design-partner consent.
The economics
What the loop is measured on
Emiteon is priced and evaluated on the numbers a fab already reports. These are design-partner targets for the first twelve months of deployment.
$26BDisplay-fab automation, process control, inspection, test and fab software market
~14%Annual growth in the segments Emiteon plays in
135%Net revenue retention target from land-and-expand
99.9%Uptime target for the fab edge runtime
Figures marked as targets are design-partner objectives, not audited results. Company operating status, customers and outcomes are [ASPIRATIONAL] until independently verified.
FAQ
Questions fabs ask first
Through three gates. Shadow mode observes and predicts with no write-back; advisory mode recommends recipe, transfer and binning moves that a process-integration or yield engineer approves; graduated autonomy releases low-risk deposition, anneal and transfer control once measured accuracy and twin validation clear the bar. High-impact decisions stay human-in-the-loop.
Reducing false calls is the wedge, not a side effect. The mura-and-defect agent is trained on true-versus-false-call labels from your own inspection history, so it separates genuine large-area non-uniformity, sub-pixel opens and shorts, particles and stains from nuisance signals, and every call is traceable to the evidence behind it.
Per-tenant isolation with recipes, panel designs and defect images scoped to your tenant, encryption in transit and at rest, SSO/RBAC, an immutable yield/quality-grade audit log and an on-prem or air-gapped option. Fleet learning shares model improvements, never your recipes.
Deposition, photolithography, encapsulation and anneal tools, Mura/AOI inspection and metrology, array and cell test, mass-transfer and repair stations, robotic handling and MES — vendor-neutral via SECS-GEM/HSMS where the tool supports it, with REST and OPC-UA bridges elsewhere.
Every pixel, perfectly uniform.
Put one agent in shadow mode
Pick the agent whose metric you already argue about in the Monday meeting.