First: what is being diagnosed?
A fab is a factory that manufactures chips on wafers. Wafers repeatedly undergo deposition, lithography, etching, cleaning, and other operations, with inspection and metrology throughout. When images show anomalies, equipment differs from its reference, or an analysis model degrades, engineers need to determine what happened, which explanations deserve investigation, and what to do next.
In this analogy, the case is an engineering problem under investigation. The engineer is responsible for it: supplying process context, setting goals, reviewing results, and deciding field actions. The AI consultation assistant organizes analysis and reasoning; scientific tools and original materials supply the evidence.
What does an engineer experience?
Consider a conceptual case: many candidate objects appear in the outer wafer region. The engineer opens the image and existing report to understand the pattern and decide which conditions to examine next. The illustration shows the intended user journey, rather than a product screenshot or completed field diagnosis.
- 01
Open the case
Images, existing analyses, and available engineering records
- 02
Frame the question
AI checks the object, scope, and intended outcome
- 03
Obtain evidence
Select checks and retrieve original values and methods
- 04
Review and assess
Check comparability and retain conflicts and gaps
- 05
Form recommendations
Explain what to check, why, and how to validate
- 06
Engineer review
Inspect original materials and decide actions and the next round
The intended deliverable is a reviewable investigation assessment that brings together findings, candidate explanations, checks or adjustment recommendations, validation steps, and evidence gaps. With sufficient support, recommendations may identify a Recipe parameter or chamber condition to examine. When the cause is unresolved, specify the next check rather than inventing an adjustment value.
What was observed? Which explanations are supported, and which remain open?
What should be checked next? Under which conditions is a check or adjustment justified?
Inspect the methods, actual results, sources, and limitations supporting each assessment.
How do expert methods guide AI while preserving judgment?
A specialist center develops stable methods: which facts to establish, when a check applies, which results are not directly comparable, and which gaps require more evidence. WaferPilot organizes reviewed methods of this kind into versioned Expertise Packages.
Methods define professional responsibilities, applicability, and evidence requirements. For each case, the Agent uses the available materials to choose actions, propose candidate explanations, review results, and update its assessment. The model can contribute knowledge and reasoning; claims about actual data still require inspectable support.
Frame the question
Which objects, quantities, and relationships must be addressed? What remains ambiguous?
Establish evidence
What was actually obtained? Is it comparable? What is missing?
Interpret with limits
Which parts can be answered? What supports the recommendation? What remains unresolved?
These steps are called rails. They belong to the professional method for a task family, allowing similar questions to share the same responsibilities. The question, materials, and selected checks are determined from the case and authorized capabilities. Different tasks can use different professional Packages.
The three rails illustrate the target design for questions about existing scientific results; they do not define every equipment-diagnosis or new-analysis execution workflow.
White-box: what can you inspect behind an assessment?
Imagine a consultation assessment accompanied by an inspectable case record: which report was used, how its measurements were defined, which method was applied, and which judgment the results support. Engineers can follow this record back to original materials instead of relying on an isolated conclusion.
- Inputs and scope
Which wafer, run, and engineering conditions were used?
- Professional method
Why was this check selected, and when does it apply?
- Actual reads and calculations
Which original values were read? Which calculations and parameters were used?
- Assessment and limitations
Which explanation is supported, and what evidence remains missing?
- Recommendation and review
What is recommended, how should it be validated, and who owns the next step?
White-box means these decisions and evidence records are visible, locatable, and reviewable, enabling errors to be found and judgments revised. Material fingerprints can verify version consistency; scientific validity requires the appropriate validation. The inspectable evidence record is distinct from the model’s internal thought process.
Explore a measurement-definition example: why are two areas not interchangeable?
The sample report records a whole-wafer inspected area and six radial areas. Both use cm², but their definitions and calculations differ. A sound explanation requires both values together with their respective definitions.
| Reported quantity | Original value | Definition |
|---|---|---|
| Whole-wafer inspected area | 678.7843 cm² | Inspected site count multiplied by die area. |
| Sum of six radial areas | 676.154504 cm² | Sum of the recorded intersections between the field-of-view (FOV) union and the radial bands, using a raster approximation. |
This supports an explanation that the definitions differ and the denominators are not directly interchangeable. The difference alone does not establish an algorithm error or equipment fault, nor attribute the entire difference to a single geometric factor. The assistant should explain the known definitions and what remains unverified.
The 561 inspected sites describe a site population. For objects/cm², the denominator is the applicable area; site count and area must not be interchanged. Annotated objects are also not automatically confirmed defects.
Map the familiar roles to platform components.
These names help explain responsibilities. Understand what each layer does first, then learn its name in the system.
| Diagnostic-center role | Platform counterpart | Responsibility |
|---|---|---|
| Case under investigation | Images and engineering context | Supply observations, actual conditions, and an investigation goal. |
| Specialist methods | Expertise Package | Preserve reviewed methods, applicability, exceptions, and evidence requirements. |
| Consultation responsibilities | Rails | Frame questions, establish evidence, and explain supported conclusions and limits. |
| AI consultation assistant | Grounding Agent / LLM | Reason within professional methods and current materials, choose actions, and propose judgments. |
| Consultation coordination | Kernel / Session | Manage permitted actions, state, accepted judgments, execution records, and recovery. |
| Testing tools | Pipeline / Methods | Execute scientific analyses or authorized reads and calculations, recording inputs and results. |
| Versioned, defined reports | ELS / Scientific declarations | Verify provenance and publication versions; declare units, scope, methods, and limitations. |
| Evidence archive | Scientific RAG | Find relevant materials, retrieve original values, definitions, and sources, and organize evidence. |
| Case owner and workbench | Engineer / ImageLens / Notebook / Reports | Observe the case, review analyses and recommendations, and decide field validation and adjustments. |
Why is a complete platform and workbench needed?
An engineering investigation needs observations, calculations, and recommendations in one context. Like an imaging workstation, test records, and a consultation report, WaferPilot’s interfaces serve connected responsibilities.
ImageLens
Inspect the wafer, spatial distributions, and local images; locate associated materials and findings.
Notebook
Review analysis inputs, methods, parameters, and results; inspect the actual calculations.
Investigation assessment
Bring together candidate explanations, recommendations, validation plans, and evidence for engineer review.
The platform’s intended deliverable supports engineering judgment and follow-up action. Visualization makes observations accessible, scientific execution produces actual results, and evidence records make recommendations reviewable. Together they support an end-to-end investigation experience.
The analogy has clear boundaries.
A diagnostic assessment can recommend further checks, evidence gathering, and validation. With sufficient support, it can point to specific parameters, equipment conditions, or model stages to examine. A recommendation does not by itself establish a root cause.
Querying published results and executing a new scientific analysis are different tasks. A query plan specifies existing materials and authorized calculations; a new analysis requires its own plan, permissions, and execution workflow. Query capability must not be presented as authorization to run a new Pipeline automatically.
The intelligent diagnostic center is an analogy for how an engineering platform is organized. It does not assume an infallible model or treat record keeping as proof of correctness. Method validation, actual evidence, and engineer review remain the basis for reliable judgments.
This article explains architectural principles and intended experience. For validated capabilities and pilot scope, see the current demo and capability guide.