MAKA: A Physics-Aware Multi-Agent Decision Support Architecture for High-Precision Manufacturing

MAKA is a human-machine collaborative multi-agent decision support architecture. By separating four core modules—intent routing, quantitative analysis, knowledge retrieval, and critical validation—it addresses AI application challenges in high-precision manufacturing. In Ti-6Al-4V rotor blade processing tests, the tool execution success rate increased by 87.5 percentage points, surface precision improved 10-fold, and traceable, risk-aware decision support for high-precision manufacturing was achieved.

High-precision processing of aerospace components (e.g., Ti-6Al-4V rotor blades) requires surface deviation control within 25 microns, which demands complex compensation strategies. Current Large Language Model (LLM) applications mostly stay at the dialogue level; they cannot execute risk-constrained multi-step numerical workflows or provide auditable decision traceability, limiting their use in high-risk manufacturing scenarios.

The core design of MAKA combines LLM reasoning with domain numerical computation through agent division of labor:

1. Intent Router: Parses user queries and routes them to appropriate workflows;
2. Tool-Specific Quantitative Analyzer: Calls professional tools to perform numerical calculations (e.g., simulation, compensation algorithms) to ensure determinism and verifiability;
3. Knowledge Graph Retriever: Retrieves relevant information from knowledge bases such as material properties and process rules;
4. Critical Validator: Performs three checks—physical rationality, safety boundaries, and traceability integrity. Collaborative workflow: Engineer query → Intent routing → Quantitative analysis → Knowledge retrieval → Candidate generation → Critical validation → Human approval → Execution feedback.

The test bench targets five-axis machining of Ti-6Al-4V rotor blades and integrates three types of data: virtual machining path error field, cutting force and deflection simulation, and 3D scan detection deviation map. Deviations are decomposed into four interpretable components: path-related component, drift-related component (tool wear/machine tool drift), residual system compliance, and variability proxy—helping engineers understand the root cause of deviations.

Tool Orchestration Benchmark Test: MAKA increased tool execution success rate by 87.5 percentage points compared to the baseline in low/medium/high complexity tasks, with higher error recovery rates (automatic retry/switching schemes). Digital Twin Verification: After compensation, surface deviation was reduced from approximately 250 microns to 25 microns, a 10-fold improvement in precision. Human Feedback: Engineers recognized its clear traceability, transparent risks, and strong sense of control.

1. Specialization Over Generalization: Division of labor allows each agent to focus on specific tasks, leading to overall performance exceeding that of a single model;
2. Necessity of Explicit Validation: Three checks ensure the solution complies with physical laws, safety constraints, and is traceable;
3. Balance Between Human-Machine Collaboration: AI handles data-intensive tasks while humans retain final decision-making power, balancing efficiency and accountability.

Current Limitations: Domain specificity (requires reconfiguration to migrate to other scenarios), knowledge base dependency, insufficient real-time performance, small-scale testing (16 blades). Future Directions: Adaptive learning (improving models from production data), cross-domain migration (additive manufacturing, etc.), real-time optimization, multi-factory coordination.

MAKA proves that through architectural design, LLMs can be transformed into reliable decision tools for high-risk manufacturing scenarios. Its methodological shift (from black-box oracle to auditable system) provides design principles for industrial AI: agent division of labor, physical constraint encoding, multiple validations, human-machine collaboration, and full traceability. These principles can be extended to high-risk fields such as healthcare and autonomous driving, providing a blueprint for AI safety and accountability.