- Platform
- A Cohesive Platform Unifying Images, Investigations, Compute, Reports, and Evidence
- Methods & Engineering
- Peer-Reviewed Diagnostic Frameworks and Criteria
- Investigation Context
- Current Objectives, Telemetry, Choices, Findings, and Unresolved Gaps
- Engineering Deliverables
- Dossiers, Recommendations, Provenance Chains, and Verification Goals
This article describes public architecture principles and product design goals. The currently verified demonstration and subsequent validation scope are explained in Current Capabilities & Scope.
Engineers Need to Drive an Investigation, Not Just Receive a Single Answer
Confronted with an excursion wafer, engineers inspect morphology, confirm baselines, query tool histories, execute analytics, and plan verification runs; a conversational text reply covers only a fraction of this workflow.
WaferPilot focuses on fab engineering diagnosis and verification. Starting with images and existing analyses, it combines authorized metrology, recipes, equipment conditions, and processing records to assist EE, PE, and PIE investigations. Target deliverables include confirmed observations, candidate causes, engineering recommendations, and the analysis supporting them.
For an 'edge defect adder' excursion, engineers verify threshold consistency, contrast normal vs suspect lots, and evaluate chamber splits. The platform orchestrates these checks, detailing answered questions and evidence gaps.
Platform Overview: Method Specs, Autonomous Agents, Tools, and Workbenches
Platform modules coordinate around a unified investigation: method specs provide direction, the Agent evaluates instance facts, tools compute metrics, and the workbench enables audit and review under human authorization.
| Component | Primary Responsibility | What Engineers Can Inspect |
|---|---|---|
| Methods & Engineering | Preserving Applicability Rules, Inspection Steps, and Evidentiary Standards | Why This Pathway Was Chosen and When This Method Applies |
| Domain Agent | Interpreting Objectives, Comparing Hypotheses, and Formulating Next Steps | Selection Rationales, Alternatives, Boundary Constraints, and Open Items |
| Scientific Computing & Tools | Executing Authorized Data Audits, Statistical Analysis, and Modeling | Actual Inputs, Run Parameters, Package Versions, and Execution Logs |
| Engineering Workbench | Unifying Micrograph Review, Plan Audit, Sandboxed Execution, and Reporting | Comprehensive Investigation Context with Bidirectional Provenance Links |
Embedding Knowledge into Versioned Methods Rather Than Personal Memory
Useful experience includes how to investigate: what to confirm first, which comparisons are invalid, when to obtain more evidence, and what results only support candidate explanations. It can come from process standards, equipment documentation, standard analytical methods, field explanations, and historical cases; a complete incident library is not a prerequisite.
Historical materials are synthesized and peer-reviewed into versioned knowledge assets with defined scope and owners. AI models assist in drafting candidate specs, while human experts validate physical soundness.
Method specs and instance observations remain strictly decoupled: 'verify product and recipe comparability before chamber split analysis' is a reusable method; 'this lot ran through chamber B' is an instance fact.
For method extraction and empirical validation, continue reading Five Dimensions of Knowledge Assets and From Incidents to Method Specs.
Guided by Method Specs, Empowering Models with High-Value Engineering Judgment
Method specs define investigative questions, evidentiary criteria, and tool APIs; models evaluate options and balance parameters in context, handling routine technical choices autonomously under human delegation.
New findings may redirect investigations: examining alternative hypotheses, requesting missing traces, or concluding. Model reasoning interprets problem context while critical facts and calculations require rigorous grounding.
When objectives are ambiguous, proprietary fab context is required, or elevated permissions are needed, the platform returns control to human engineers, balancing autonomy with explicit responsibility boundaries.
ImageLens, Notebook, and Reports: Addressing Needs Across Different Investigation Stages
ImageLens anchors spatial inspection across wafers and ROIs; Notebooks execute approved plans and display reproducible code and outputs; Reports deliver structured findings, recommendations, and verification goals.
An end-to-end experience demands that workspaces share unified session identity and state. Toggling to analytics transfers selected images, references, and versions; auditing recommendations returns to exact source data.
This defines the purpose of a specialized fab workbench: high-resolution micrographs, spatial ROIs, and process logs shape physical interpretation, allowing engineers to verify the Agent handles the incident correctly.
Delivering Recommendations Alongside the Full Trajectory That Produced Them
Deliverables explicitly state: observations, isolation tier, supported hypotheses, open alternatives, and verification steps. Bounded data yields actionable recommendations; sparse data produces targeted validation requests.
Reports preserve target objects, reference baselines, method selections, parameters, outputs, and rationales, enabling engineers to audit conclusions, correct biased assumptions, and advance investigations.
White-box architecture means engineering decisions and evidentiary pathways are fully auditable, empowering verification and correction; method validity and data quality must still be empirically tested.
Decoupling Method Specs from Infrastructure Evolution
Decoupling method specs from shared execution runtimes allows the platform to onboard new analytical tools. Distinct investigations demand specialized physics while sharing data binding, plan audit, and reporting.
Wafer-defect investigations focus on inspection conditions, distributions, and processing history; model-degradation investigations focus on evaluation sets, labels, input changes, and model versions. They can reuse some investigation and review mechanisms, but methods, input structures, and acceptance criteria must be designed separately. Automatic training and deployment also require corresponding capabilities and authorization.
Extensibility is an architectural cornerstone. Entering a new process domain requires spec authoring, tool integration, and empirical verification; cross-domain efficacy must be demonstrated by reproducible findings.
How to Continue Reading This Architecture Series
To explore technical mechanisms first, read Agent Architecture; to understand how knowledge becomes a digital asset, continue with Five Dimensions of Knowledge Assets and Case to Method; for reliability and validation, read White-Box Evidence; for user experience and workflow, read Full Workspace.
Deployment topologies and ecosystem integration have dedicated guides: Staying Within Customer Environments: Deployment and Collaboration. Current demonstrator capabilities and validation scopes are detailed in Current Demonstrator and Scope of Capabilities, distinguishing architecture goals, interaction illustrations, and verified deliverables.