Blog

Fabs and autonomy, written plainly.

Short pieces from the Waferira team on process control, defect classification, virtual metrology and what it takes to let software write a recipe. Four so far. Each one is drawn from the thinking in our own planning documents, not from customer work, because there is none yet.

Read the latest About Waferira

4 postsWaferira team2026-10-08

Latest

Rich data, reactive control

Fabs collect more data per wafer than any factory on earth, and still find excursions after the lots that ran through them are gone.

2026-10-08 · 5 min read

2026-10-08 · 5 min read · Waferira team

Rich data, reactive control

Fabs collect more data per wafer than any factory on earth, and still find excursions after the lots that ran through them are gone.

A front-end fab is the most capital-intensive, precise and complex factory humanity has built. A wafer passes 700 to 1,500 process steps over months. Every layer has to hit nanometer overlay, critical dimension, film thickness and uniformity across thousands of tools, fully traceable. And a few stray nanometer-scale defects can destroy a lot worth hundreds of thousands to millions of dollars.

The data is already there: inline inspection imagery, metrology, tool sensors, FDC, yield. What is missing is the loop between seeing a problem and acting on it. Inspection tools generate images engineers review by hand. APC and FDC run in silos, so an etch or deposition excursion surfaces after the damage. Dispatch is rule-based and blind to live yield risk. Metrology is sampled, so most wafers and steps go unmeasured.

Each gap is survivable alone. Together they cap yield, stretch cycle time and burn a process and yield engineering workforce that is already too thin. Waferira's premise is that the fab has the sensors, the metrology and the traceability; what it lacks is autonomy that turns all of that into action on the next wafer.

2026-10-08 · 4 min read · Waferira team

Autonomy is graduated, not switched on

No fab goes from manual APC to unattended control in one step. The level has to be explicit, auditable and reversible.

Design partners start in shadow mode. Waferira connects to the wedge module, watches, measures the baseline for yield, excursions, cycle time and engineer load, and writes nothing. In assist mode it recommends recipe moves, defect dispositions and dispatch, and a process engineer approves each one. Bounded autonomy follows for low-risk APC, defect and metrology loops, inside the qualified recipe window, with exceptions escalated and the OCAP linked.

Every agent action carries its source signals, the procedures it retrieved with citations, the predicted impact, the allowed range, the rollback condition, the responsible approver and the lot scope. That record is what makes the system legible to the fab, to yield engineering and to IT and OT security, and it is what makes trust possible before autonomy.

2026-10-08 · 4 min read · Waferira team

Hit yield before the lot runs

Fabs commit lots and node ramps blind. The twin changes the order of operations.

New recipes, process integrations, WIP moves and dispatch changes are tried in production, where a bad move scraps lots and burns cycle time. There has been no as-run, physics- and flow-accurate model of the whole fab to test a candidate process, metrology or dispatch decision before execution, so engineers rely on craft, spreadsheets and costly split-lot experiments.

Fabtwin fuses process and flow into an Omniverse-based twin that simulates candidate recipes, correctables, WIP and dispatch moves, as-run against as-designed, and returns the predicted yield, uniformity and cycle time before the lot runs. Every Lithera, Vmetra, Scanira and Dispira action is validated there first, and every simulated-versus-qualified decision is stored as process-qualification memory.

Physics-ML surrogates, thousands of daily simulations and real-time twin rendering are GPU-bound workloads. Running them fast enough for autonomy is the reason the twin is built on accelerated computing.

2026-10-08 · 3 min read · Waferira team

The craft does not scale. The data does.

The people who could close the loop are scarce and getting scarcer, right as reshoring and advanced nodes add fabs and tighten tolerances.

Talk to a fab director and the same losses come up: an overlay drift that reached the control limit a lot too late, a particle excursion caught at final test, a defect-review backlog nobody can clear, a missed ramp because the handful of engineers who understand a process could not keep up with the data. The craft that prevents these lives in a few people's heads.

Waferira captures that craft as a model: the full chain of process state, prediction, approved action and later yield outcome, per lot. Every supervised correction trains the system, and rare defect and excursion learnings can transfer across similar process families where contracts allow, strictly IP-isolated. The fab runs at expert attention on every shift, and the record of how is written down.

Run timeline · scenario

The scenario the posts refer to

Every Waferira run is an ordered, inspectable sequence. This is run_7b2e9c on Module M2 · 300 mm · metal-1 · lot L-48213. Scenario completes in 10 steps with one approval gate. Overlay and CD predicted inside spec, one reticle-repeater signature caught and held for review, dispatch re-optimised, lot released with full genealogy.

run_7b2e9c Module M2 · 300 mm · metal-1 · lot L-48213 illustrative scenario SUCCEEDED
  1. 01 ingest.lot SUCCEEDED 0.7 s

    Fab Orchestrator · Pulled lot L-48213 from fab MES: 25 wafers, 300 mm, metal-1 layer, route, reticle, overlay and CD spec and the standing cycle-time target. Locked the target envelope for the run.

  2. 02 twin.simulate SUCCEEDED 38 s

    Fab Twin · Simulated 64 candidate correctable, recipe and dispatch sets on the as-run fab twin and ranked them on predicted overlay residual, CD uniformity, excursion risk and cycle time.

  3. 03 litho.correctables SUCCEEDED 1 m 20 s

    Litho-and-Overlay · Per-field overlay correctables and a focus-dose trim written to scanner SCN-07 over SECS/GEM after the previous lot's residual trended toward the control limit. Predicted overlay residual back inside spec.

  4. 04 etch.uniformity SUCCEEDED 58 s

    Etch-and-Depo · Chamber B on etcher ETCH-12 showed an edge-fast uniformity drift in FDC. Pressure and RF trims written within the qualified recipe window; predicted CD uniformity recovered.

  5. 05 cmp.removal SUCCEEDED 46 s

    CMP-and-Implant · Downforce and platen-speed trim on CMP-03 against the virtual removal-rate estimate to hold post-CMP thickness and dishing inside the window.

  6. 06 inspect.classify SUCCEEDED continuous

    Defect-and-Inspect · Inline optical, e-beam review and SEM imagery fused at the tool edge. 1,412 events classified, most dispositioned nuisance; one reticle-repeater signature on three dies flagged for review.

  7. 07 vm.score SUCCEEDED 2.1 s

    Metrology-and-Virtual · Virtual metrology scored every wafer at post-etch and post-CMP from tool-sensor and FDC signatures. Two wafers routed to physical metrology on predicted variance; the rest sampled at plan.

  8. 08 approve.human APPROVAL 1 m 04 s

    Fab Orchestrator · The reticle-repeater signature called for a reticle inspection hold and a recipe move outside the qualified window. Held for the module process engineer and approved at 03:14:27.

  9. 09 dispatch.route SUCCEEDED 4.3 s

    Yield-and-Dispatch · Dispatch re-optimised across 1,840 lots in WIP under reticle, chamber and qualification constraints. The lot routed to the qualified chamber with the shortest predicted queue; AMHS move issued to the stocker.

  10. 10 lot.release SUCCEEDED 1 m 05 s

    Quality-and-Conformance · Lot L-48213 moved on: genealogy written (recipe moves, correctables, FDC signatures, VM scores, defect map, approvals) and checked against the layer spec.

Autonomy

Four levels, set per agent and per tool

A fab does not go from manual APC to unattended control in one step. Waferira makes the level explicit, auditable and reversible at any time. The first release runs shadow mode, then assist mode, then graduated autonomy for low-risk APC, defect and metrology loops.

Autonomy levels and the human role at each
LevelWhat the agent doesWhat the human doesWhere it sits in a pilot
L1 · AdvisoryRecommends recipe moves, dispositions and dispatch, with citationsEnters every change manuallyShadow mode, first on the design-partner cell
L2 · SupervisedProposes a recipe write or hold; it executes on approvalApproves each move in the review consoleAssist mode, where accuracy and ROI are proven
L3 · BoundedWrites inside the qualified recipe window, tool-state checks and OCAP linkageApproves holds, out-of-window moves and lot releaseGraduated autonomy for low-risk APC, defect and metrology loops
L4 · Qualified loopsRuns a qualified loop without prompting, inside its envelopeSets the envelope; reviews the shift recordPlanned; reserved for loops the fab has qualified
Accelerated computing

Physical AI, at the tool

Computational lithography, defect classification on e-beam and SEM imagery, per-wafer virtual metrology and a full-fab twin are among the most compute-intensive workloads in manufacturing. Waferira keeps IP-sensitive inference and feature storage at the fab edge and runs training, simulation and optimisation on GPUs the fab controls.

Metropolis · DeepStream · Holoscan · TensorRT

Inspection vision

Inline optical, e-beam and SEM imagery and sensor fusion on fab-edge GPUs. 10 to 50 inspection streams per pilot module and sub-200 ms priority defect triage are the design targets.

CUDA · RAPIDS · Triton

Virtual metrology and FDC

High-volume multivariate tool-sensor and FDC time-series scoring on every wafer, 1 to 5 seconds per lot-step in the design, with nightly retraining as recipes and chambers drift.

DGX or HGX · CUDA · cuDNN

Training

Defect, virtual-metrology and process-response models retrained per process family; 500 to 2,000 GPU-hours per family per quarter is the design target. Dedicated capacity is planned.

Omniverse · OVX

Fab twin

As-run process and flow simulation, 1,000 to 10,000 candidate recipe, metrology and dispatch scenarios per fab per day in the design. OVX access is planned.

Cosmos · Omniverse Replicator

Synthetic rare events

Rare defect, particle, overlay-drift, CMP-dishing and film-uniformity variants generated and validated against real distributions before promotion. Planned.

cuOpt · RAPIDS

Dispatch optimisation

Constrained WIP dispatch re-optimised every 1 to 5 minutes across reticle, chamber, route, tool-state and cycle-time constraints.

Jetson Orin · Isaac · TensorRT

AMHS and handling

Local FOUP and robot perception and exception detection at the fab edge. Jetson Thor is planned for higher sensor density.

NIM · NeMo · Retriever · Guardrails · AI Enterprise

Process reasoning

Tenant-isolated process Q&A and OCAP reasoning over recipes, SEMI procedures and audit logs, with citations enforced on every answer.

Who it is for

Built for the people who run the fab

We have no customer quotes to show yet and will not invent any. These are the three roles in a fab we build for, and the problems they bring to the first meeting.

Fab, Operations or Yield Director

Owns yield, throughput, cycle time, excursion response and cost per wafer.

What this person is dealing with today

  • Reactive process control that lets overlay, uniformity and CMP excursions destroy yield
  • Multi-million-dollar lot scrap from undetected excursions
  • A severe, scarce process and yield engineering shortage that caps the ramp

Process Integration or Module Engineer

Owns litho, etch, deposition, CMP and implant recipes, process windows and APC.

What this person is dealing with today

  • APC, defect, metrology and MES siloed and reactive across the fab
  • Correctables that lag because metrology is sampled
  • Rule-based dispatch and WIP management that inflate cycle time

Yield, Defect or Metrology Engineer

Owns defect classification, metrology, virtual metrology, yield analytics and wafer and lot genealogy.

What this person is dealing with today

  • Manual, engineer-bound defect classification and SEM review drowning in data
  • Metrology bottlenecks and sampling gaps that miss yield-limiting excursions
  • Genealogy stitched by hand across tool, metrology and MES data
FAQ

Straight answers

The questions fab directors and process engineers ask in the first meeting.

Yes, over SECS/GEM, but only inside the qualified recipe window for that tool, with per-parameter limits and tool-state checks, and only at the autonomy level your fab has set. Level 1 is advisory: Waferira recommends and an engineer enters everything. Design partners start there, in shadow mode.

Get started

Talk to the people who wrote this

Start with one process module or tool fleet and one measurable baseline. A paid, IP-isolated pilot on defect classification, virtual-metrology-driven overlay and CD control or uniformity APC shows the number before you commit further.

Book a module review See pricing