[Paper] Physics-Grounded Multi-Agent Architecture for Traceable, Risk-Aware Human-AI Decision Support in Manufacturing

Source: arXiv - 2605.04003v1

Overview

The paper introduces MAKA (Multi‑Agent Knowledge Analysis), a human‑in‑the‑loop decision‑support system that couples large language models (LLMs) with physics‑based simulations, knowledge‑graph retrieval, and a verification “critic” to guide high‑precision CNC machining of aerospace rotor blades. By enforcing physical plausibility, safety bounds, and full provenance, MAKA makes AI‑generated recommendations traceable and risk‑aware—something traditional LLM assistants can’t guarantee in safety‑critical manufacturing.

Key Contributions

• Hybrid multi‑agent architecture that separates intent routing, quantitative tool execution, knowledge‑graph lookup, and a physics‑grounded verification layer.
• Traceability & provenance: every recommendation is linked to raw inspection data, simulation outputs, and the reasoning steps that produced it.
• Risk‑aware verification critic that checks physical feasibility, safety limits, and completeness before surfacing suggestions to the human operator.
• Empirical validation on a Ti‑6Al‑4V rotor‑blade testbed, integrating virtual‑machining error fields, cutting‑force/deflection simulations, and 3‑D scan deviation maps from 16 real blades.
• Performance boost: up to 87.5 pp improvement in successful multi‑step tool execution compared with a naïve single‑LLM interaction model.
• Digital‑twin what‑if studies showing MAKA can generate compensation plans that shrink predicted surface deviation from ~10⁻² in to ~±10⁻³ in across most of the blade.

Methodology

1. Intent Routing Layer – The user’s high‑level goal (e.g., “reduce blade surface error”) is parsed by an LLM that decides which downstream agents to invoke.
2. Tools‑Only Quantitative Agents – Specialized agents run deterministic, physics‑based tools:
   • Virtual‑machining path‑tracking to compute error fields,
   • Cutting‑force & deflection simulators for tool‑workpiece interaction,
   • 3‑D inspection data retrieval from a knowledge graph of scanned blades.
3. Knowledge‑Graph Retrieval – A graph database stores historical machining parameters, material properties, and inspection outcomes. Agents query it to fetch context‑specific priors (e.g., wear trends).
4. Critic‑Based Verification – Before any recommendation reaches the human, a “critic” agent checks:
   • Physical plausibility (does the suggested compensation obey material limits and kinematics?),
   • Safety bounds (are cutting forces within allowable ranges?), and
   • Provenance completeness (are all data sources and reasoning steps logged?).
5. Human‑in‑the‑Loop Approval – The operator sees a concise, traceable report and can accept, modify, or reject the AI‑generated compensation plan.

The architecture is implemented as a sequence of stateless or stateful agents, allowing 1‑step up to ≥3‑step workflows. Each step’s output becomes input for the next, enabling complex, multi‑stage reasoning that a single LLM prompt cannot reliably achieve.

Results & Findings

The digital‑twin simulations demonstrated that MAKA can automatically generate compensation vectors that bring the blade’s predicted geometry well within tolerance, providing a “pre‑deployment verification signal” that engineers can trust before committing to actual machining.

Practical Implications

• Reduced scrap and rework: More accurate compensation means fewer out‑of‑tolerance parts, directly cutting material waste and machining time.
• Accelerated onboarding: New engineers can rely on the traceable AI reports to understand why a particular compensation is suggested, shortening the learning curve.
• Regulatory compliance: Full provenance satisfies audit requirements common in aerospace and defense manufacturing (e.g., AS9100, ISO‑9001).
• Scalable AI integration: The modular agent design lets factories plug in existing simulation tools or proprietary knowledge bases without rewriting the whole system.
• Risk‑aware automation: By enforcing safety bounds before any machine command is issued, MAKA enables higher levels of autonomy while keeping human oversight where it matters most.

Limitations & Future Work

• Domain specificity – The current implementation is tuned to Ti‑6Al‑4V rotor‑blade machining; transferring to other materials or part geometries will require re‑training or new simulation models.
• Simulation fidelity – Accuracy hinges on the quality of the virtual‑machining and force‑deflection models; any modeling error propagates to the compensation plan.
• Scalability of knowledge graph – As the repository grows, query latency could become a bottleneck; future work should explore graph‑partitioning or caching strategies.
• Human‑critic interaction – While the critic automates many checks, nuanced engineering judgment (e.g., trade‑offs between surface finish and tool wear) still needs richer UI support.
• Extending to closed‑loop control – The authors plan to integrate real‑time sensor feedback so that MAKA can adapt compensation on‑the‑fly during machining, moving from decision‑support to autonomous control.

Authors

• Danny Hoang
• Ryan Matthiessen
• Christopher Miller
• Nasir Mannan
• Ruby ElKharboutly
• David Gorsich
• Matthew P. Castanier
• Farhad Imani

Paper Information

• arXiv ID: 2605.04003v1
• Categories: cs.MA, cs.AI, cs.IR
• Published: May 5, 2026
• PDF: Download PDF