System design for physical AI: from a defect image to the chamber that drifted

Posted on Nov 3 • Originally published at codeatoms.ai

By Muhammad Umar. This is the engineering summary of an open reference architecture. The paper, its object model as JSON and the model register are free to reuse under CC BY 4.0: https://codeatoms.ai/wafer-defect-inspection-us/. DOI 10.5281/zenodo.23889770. The operator is described by class, never by name.

A high volume wafer fab in the United States images wafers at many process steps with optical inspection and electron microscope review tools. When a wafer map shows a ring signature, the yield engineer needs three answers at once: which chamber drifted, when it started, and which lots it touched before anyone noticed. The manufacturing execution system knows the lot and the hold, the yield management system knows the signature, the inspection tools' own databases hold the images, the equipment data collection system holds the chamber telemetry and the statistical process control system holds the limits. Today an engineer does the join by reading five screens while material keeps moving through the drifting chamber.

The design is a projection over the fab's systems of record, which stay untouched: never replaced, never modified, never loaded. Eight sources feed it, the five systems of record above plus the tool interfaces and telemetry streams at the edge, all read only. One integration adapter family carries every record in over SECS/GEM, GEM300 and EDA interfaces, chosen tool by tool by which revisions each etch module tool implements.

The tier stamps every record with its provenance and timestamp, constructs the event log in the pipeline rather than trusting what arrives, and proves the key joins on one real week of production data before phase one is judged. The only write anywhere is an approved hold or disposition through the manufacturing execution system interface, and a hold placed or released without a recorded named approval is refused in code and audited per hold.

Apache Kafka in KRaft mode orders events per wafer and per tool. Adapters deliver at least once and idempotently, so a replay cannot double count a defect or a confirmed class, and the backbone absorbs bursts when a review tool emits images faster than the edge tier classifies them. Linuxptp against an Open Time Card grandmaster class clock puts an FDC trace, an inspection run and a wafer map signature from the same excursion on one clock. Time series and telemetry sit in InfluxDB 3 Core, containers run on K3s, artifacts live in an air gapped Harbor registry, identity in Keycloak, and Prometheus, Grafana and Loki watch the running system. Nothing on the backbone crosses the fab perimeter.

The object model holds sixteen objects with twelve typed links: wafer, lot, process step, tool, chamber, recipe, inspection run, defect, defect class, wafer map signature, excursion, hold and corrective action, plus the defect review engineer, the yield engineer and the process owner. Each is anchored in the system that owns its truth. One query starts at a defect image classed as a pattern defect, walks to its wafer, to the ring signature that wafer carries, to its lot, back through the inspection run to the tool and recipe, to the chamber whose running hours since maintenance and FDC trace explain the drift, then to the excursion, the hold and the corrective action that closes it. It is one of our industrial ontology examples, published as JSON.

A CodeNinja on device AI stack unit sits beside each inspection tool group: an octa core edge SoC class with a 6 TOPS NPU, fanless in a sealed smooth wall cleanroom enclosure with a power over Ethernet front end, because a fan sheds and moves particles. An off the shelf GPU edge module was set aside, since its 7 to 25 W sits badly in a sealed fanless enclosure. The unit runs an unsupervised anomaly model fitted on the fab's own normal images, picked against a classifier trained on labelled images, which locks the estate to one defect mix and becomes a standing retraining burden.

Confirmed classes and root causes become memory the next decision reads, and nothing is trained into weights. The shelf load figures were measured for detectors, not this model, so a 60 percent headroom gate governs until the real load is measured, and no accuracy is quoted before the phase one bench run on the fab's own images. No new cameras are added; the tools' own images are consumed.

Three agents draft. The classification agent proposes a defect class for the defect review engineer to confirm, the excursion agent traces a signature that crosses a control limit to a chamber and drafts a hold for the yield engineer, and the disposition agent drafts release, rework or scrap for the process owner. Reasoning is intended for GLM 5.3 open weights, self hosted inside the fab, and chapter 3 places its 753 billion parameters at FP8 on one full node of the 141 GB HBM GPU class, eight GPUs per node.

The register itself holds no verified weights yet: no open weight model survived the run's live checks, so the footprint stays formally unsized until a verified row is filed with its parameter count, precision and license terms. Because a fab's process data is controlled technology, every supporting engineer is screened under the EAR deemed export rule (15 CFR 734.13) before access, and the check is carried as a named approval.

Appendix A prices the one node chapter 3 names at about 576,000 US dollars to own over three years (551,000 to 623,000), with power at the United States average industrial price. Renting the same eight cards on the deepest three year commitment in AWS US East (N. Virginia) is about 625,000, so ownership costs about nine tenths of it, and about 1.66 million on demand, against which ownership costs about a third. Renting is a price reference only, because every rented option would carry fab data out of the fab.

The rollout is two phases and 18 scope items, 17 for one etch module over one quarter and 1 for the extension gates. It is measured as time from excursion onset to hold against a nominal of 4.0 hours, exercised by one etch chamber drifting and printing a ring signature on wafer maps.

Made with CodeNinja Praxis, the CodeNinja Atoms platform for designing physical AI systems. The object model imports into Hyper Ontology. Paper, ontology and register: https://codeatoms.ai/wafer-defect-inspection-us/

One of the CodeNinja Atoms worked examples of physical AI system design: designs that sense an operation, reason over typed objects and act through a named person, each with its compute and cost.