Topics in AI Systems: Part III — GNN+LLM Architecture and Implementation Patterns

In Part I of this series, we explored why fusing Graph Neural Networks (GNNs) and Large Language Models (LLMs) is essential for solving complex problems in digital industries. In Part II, we reviewed what major model families developers should choose from when building fusion solutions. In this final article, we move from why and what to how. We introduce practical reference architecture and implementation patterns for building GNN+LLM systems in real industrial environments (see Figure 1). The key message of this article is simple: GNN+LLM fusion is not a single model; it is a system architecture.

Figure 1: High-level GNN+LLM reference architecture illustrating how fusion helps human workers.

I. From AI Models to AI Systems

In research papers, fusion often appears as a clever model design. In production, fusion becomes a multi-layer platform integrating data pipelines, knowledge graphs, analytics, retrieval, reasoning, orchestration, and user workflows. Industrial AI succeeds or fails based on system architecture, not model selection.

In this article, we introduce a reference architecture that aligns with the digital thread and digital twin vision used across aerospace, automotive, industrial machinery, and electronics. A practical fusion platform (i.e., reference system architecture for Industrial GNN+LLM fusion) typically consists of seven layers:

Layer	Purpose
1. Data Sources	Engineering, MCAD/ECAD, lifecycle operational data, etc.
2. Knowledge Graph	Unified digital thread representation
3. GNN	Structural reasoning and analytics
4. Retrieval	Grounded knowledge access with the latest data
5. LLM	Reasoning, explanation, and interaction
6. Orchestration, Traceability	Tools, governance, and automation
7. Application	Engineering copilots and workflow automation

Each layer plays a distinct role in connecting system structure with engineering knowledge to assist the human worker.

Layer 1 — Data Sources: The Digital Thread

Fusion begins with the digital thread (i.e., the connected lifecycle data of a product or production system). Typical industrial data sources include:

• Product structures and BOMs
• Requirements and change requests
• MCAD/ECAD and simulation results
• Product & Manufacturing Information (PMI)
• Manufacturing planning and execution data
• Supplier and supply chain relationships
• Service and maintenance records
• Engineering documentation and standards

When these data sources remain siloed, AI systems struggle to produce trustworthy insights for humans. Fusion requires bringing these sources together into a unified representation.

Layer 2 — Knowledge Graph: The Bridge Between Worlds

The knowledge graph is the central meeting point of GNNs and LLMs. It connects:

• Components
• Requirements
• Simulations
• Suppliers
• Documents
• Manufacturing processes
• Service events

This graph becomes the structural backbone of the digital twin: LLMs understand language, GNNs understand relationships, and the knowledge graph connects both.

Layer 3 — The GNN Layer: Structural Reasoning

In production systems, GNNs do not generate text; they generate relationships and insights about system structure. Typical industrial GNN tasks include:

• Design change impact analysis
• Failure propagation prediction
• Component similarity detection
• Supply chain risk analysis
• Anomaly detection across fleets or factories
• Relationship discovery in product and lifecycle data

These outputs provide structured context that can guide retrieval and reasoning.

Layer 4 — Retrieval: Where Fusion Becomes Real

Retrieval Augmented Generation (RAG) becomes significantly more powerful when guided by graph insights. A typical fusion workflow:

1. User asks a question about a subsystem.
2. The GNN identifies relevant components and relationships.
3. Retrieval systems fetch documents and records linked to those nodes.
4. The LLM produces a grounded response based on verified sources and latest information.

This combination dramatically improves relevance, traceability, and trust.

Layer 5 — The LLM Layer: Reasoning and Interaction

The LLM acts as the natural interface between engineers and complex data. Typical roles:

• Explaining analysis results
• Summarizing engineering knowledge
• Answering technical questions
• Generating reports and workflows
• Assisting with decision making

In fusion systems, the LLM becomes the communication and reasoning layer, not the sole source of intelligence.

Layer 6 — Orchestration, Governance, and Traceability

Enterprise deployment requires more than models. Critical capabilities include:

• API orchestration and workflow automation
• Tool integration with engineering platforms
• Security and trustworthy access control
• Auditability and traceability of outputs
• Human-in-the-loop validation

In regulated industries, traceability and trust are not optional. Fusion architecture must support explainable and verifiable outputs and ensure that final decisions are made by a human.

Layer 7 — Application Layer: Engineering Copilots and Automation

When the lower layers are in place, organizations can build new classes of applications:

• Engineering copilots
• Manufacturing assistants
• Service intelligence tools
• Supply chain risk advisors
• Automated reporting and analysis workflows
• …

These applications turn AI from a research capability into daily productivity tools for workers.

Figure 2: Depiction of seven layers in the reference architecture for fusion of GNN+LLM.

II. Implementation Patterns in Digital Industries

Several recurring patterns are emerging across organizations.

Pattern 1 — Engineering Copilot

Goal: Assist engineers with system understanding.

• GNN identifies relevant subsystems
• Retrieval gathers related documentation
• LLM provides grounded explanations and answers

Benefits:

• Requirement traceability
• Design rationale exploration
• Certification preparation

Pattern 2 — Change Impact Analysis

Goal: Understand consequences of design changes.

• Engineering change proposed
• GNN predicts affected components and processes
• LLM generates impact reports and recommendations

Benefits:

• Faster change cycles
• Reduced downstream risk
• Improved cross-team communication

Pattern 3 — Manufacturing Optimization Assistant

Goal: Improve production efficiency.

• Production line modeled as a graph
• Temporal GNN detects bottlenecks or anomalies
• LLM explains root causes and suggests actions

Benefits:

• Reduced downtime
• Faster troubleshooting
• Knowledge-capture from operations data

Pattern 4 — Service Intelligence Assistant

Goal: Improve maintenance and fleet performance.

• Fleet telemetry feeds a dynamic graph
• Temporal GNN detects patterns and anomalies
• LLM supports troubleshooting and root-cause analysis

Benefits:

• Predictive maintenance
• Faster field service support
• Reduced lifecycle costs

III. Adoption Roadmap

Our organization can evolve these patterns through multiple phases:

Phase	Focus
Phase 1	Build knowledge graph and RAG foundation
Phase 2	Add GNN analytics for structural insights
Phase 3	Introduce engineering copilots via LLM
Phase 4	Automate workflows and decision support

Starting small and expanding iteratively reduces risk and accelerates value creation.

IV. Common Pitfalls and Lessons Learned

Successful programs often share similar lessons:

• Avoid starting with complex agents before data readiness.
• Invest early in knowledge graph and retrieval foundations.
• Prioritize governance and traceability from the beginning.
• Focus on high-value use cases and trustworthiness before scaling.

We emphasize again that fusion succeeds when it is treated as a system rather than mere selection of AI models.

V. Conclusions

Across this three-part series of articles on the fusion of GNNs and LLMs, specific to Digital Industries, we explored:

• Part I: Why combining structural reasoning and language understanding is essential;
• Part II: What major GNN and LLM model families are available to developers; and
• Part III: How practical architecture and proper implementation patterns could lead to successful industrial deployments.

The convergence of GNNs and LLMs enables a new class of AI systems that can understand both the structure of engineered systems and the knowledge surrounding them. Such a fusion brings AI closer to the daily workflows of engineers, planners, and service teams, supporting better decisions by workers across the entire product lifecycle.

As digital threads and digital twins continue to mature, the integration of GNNs and LLMs will play a central role in transforming how complex systems are designed, built, and maintained. Hopefully, this three-part series will play a small role in such a transformation of Digital Industries. Note: This document was prepared with the help of AI, and An important correction to the original draft was made by my colleague Matthew McMinn.