GENAI WORKBENCH: AI-Assisted Analysis and Synthesis of Engineering Systems from Multimodal Engineering Data

This paper introduces the GenAI Workbench, a conceptual Model-Based Systems Engineering framework that integrates multimodal data and generative AI to bridge the gap between detailed component design and holistic system architecture by automatically extracting requirements and generating system structures from source documents.

The Gist

Imagine you are building a massive, complex Lego spaceship.

In the current world of engineering, the process is a bit chaotic. You have one team writing the instructions (the Requirements): "The wings must be strong enough to withstand wind." You have another team building the actual bricks (the CAD/Geometry): "Here is the 3D model of the wing." And you have a third team drawing the map of how everything connects (the System Architecture): "The wing connects to the fuselage here."

The problem? These three teams are working in different rooms, speaking different languages, and using different notebooks. If the instructions change, the brick builders might not know. If the map changes, the instructions might become outdated. This creates "silos" where mistakes hide until the very end, leading to expensive failures.

Enter the "GenAI Workbench."

Think of this paper as a blueprint for a super-smart, magical project manager that sits right in the middle of your workspace. It doesn't just store files; it understands them and connects them automatically. Here is how it works, broken down into simple concepts:

1. The "Universal Translator" (The Digital Thread)

Right now, a text document, a 3D drawing, and a spreadsheet don't "talk" to each other. The GenAI Workbench acts like a universal translator.

• The Analogy: Imagine every single part of your spaceship has a unique barcode (a Unique ID). The Workbench scans the text document, sees the barcode, scans the 3D model, sees the same barcode, and scans the connection map, and sees it again.
• The Result: It creates a "Digital Thread"—a single, unbreakable string of information that ties the idea (text) to the shape (geometry) to the plan (graph). If you change the text, the system instantly knows which 3D shape and which connection map are affected.

2. The "AI Co-Pilot" (How it Works)

The paper proposes a workflow where an AI acts as a helpful assistant to the engineer, doing the boring, repetitive work so the human can focus on the big picture.

• Step 1: Reading the Manual (Document Ingestion)

  • Current Way: An engineer reads a 100-page PDF and manually types out 50 requirements into a database.
  • The Workbench Way: You drop the PDF into the system. The AI (a "Large Language Model") reads it, finds the important rules, and automatically writes them down in the database. It's like having a robot secretary who never misses a detail.

• Step 2: Drawing the Blueprint (Architecture Synthesis)

  • Current Way: The engineer spends weeks figuring out how the engine, wings, and computer connect, drawing a complex chart by hand.
  • The Workbench Way: The AI looks at the requirements it just read and says, "Based on these rules, here is a draft map of how the parts should connect." It generates a Design Structure Matrix (a fancy grid showing who talks to whom).
  • The Catch: The AI isn't perfect. It might guess wrong. That's why...

• Step 3: The Human Check (Human-in-the-Loop)

  • The engineer looks at the AI's draft. "Ah, you forgot the solar panels!" or "No, the battery doesn't connect to the antenna." The engineer fixes it.
  • The Magic: Once the human approves it, the system locks that decision in. The AI learns from the correction, and the "Digital Thread" is updated instantly.

3. The "Reality Check" (Verification)

In traditional engineering, you might build the whole ship and then realize the wings are too heavy for the engine.
The GenAI Workbench acts like a continuous safety inspector.

• Because it links the text (the rule) to the 3D shape (the part), it can constantly check: "Does this specific wing shape satisfy the weight rule in the text?"
• If you change the wing shape in the 3D model, the system immediately checks if it still breaks the rules. It catches errors in real-time, not at the end.

4. Why This Matters (The "CubeSat" Example)

The paper uses the example of building a CubeSat (a tiny, cheap satellite).

• Before: A small team of students spends months trying to make sure their battery, camera, and radio all fit and work together. They miss a connection, and the satellite fails in space.
• With GenAI Workbench: The AI helps them draft the whole plan in hours. It automatically checks if the camera fits the power budget. It ensures that if the launch rocket changes its rules, the students know immediately which parts of their satellite need to change.

The Big Picture

This paper isn't selling a product you can buy today. It's a proof-of-concept. It's saying: "We have the tools (AI, 3D modeling, databases) to build a system where engineering is seamless, connected, and intelligent. Let's stop working in silos and start building a 'Digital Thread' that holds everything together."

It's the difference between building a house with three separate teams who never talk to each other, versus having one super-intelligent architect who ensures the blueprint, the bricks, and the plumbing are perfectly in sync from day one.

Technical Summary

Here is a detailed technical summary of the paper "GenAI Workbench: AI-Assisted Analysis and Synthesis of Engineering Systems from Multimodal Engineering Data."

1. Problem Statement

Modern engineering design is currently fragmented. While discipline-specific tools (CAD, CAM, CAE) excel at detailed component design, they lack integrated Model-Based Systems Engineering (MBSE) frameworks. This creates "information silos" where:

• System-level requirements and architectures are managed separately from detailed physical designs.
• The "digital thread" (the continuous flow of data across the product lifecycle) is broken.
• Verification and validation (V&V) are complicated, leading to increased integration risks and integration failures.
• Current MBSE tools are often manual, siloed, and do not leverage Artificial Intelligence (AI) to bridge the gap between unstructured data (documents) and structured data (models).

2. Methodology: The GenAI Workbench

The authors propose the GenAI Workbench, a conceptual framework and integrated software environment designed to embed AI-driven systems engineering directly into the designer's workflow. It is not a commercial product but a methodological proof-of-concept built on an open-source Product Lifecycle Management (PLM) platform.

Core Technical Architecture

The system integrates three distinct data modalities using a Unified Digital Thread anchored by Persistent Unique Identifiers (UIDs):

1. Document (Semantic): Textual requirements, standards, and design briefs.
2. Geometry (Physical): CAD models (B-rep, STEP formats) representing physical components.
3. Graph (Relational): System architecture graphs (e.g., Design Structure Matrices) representing interactions and dependencies.

The architecture relies on a Systems Hub (central repository), a DocumentLabeler (semantic engine), and an Architecture Model (visual canvas).

The AI-Assisted Workflow

The workbench operates through a sequential, human-in-the-loop workflow:

1. Document Ingestion:
   • Source documents (PDF/Word) are processed using OCR and Large Language Models (LLMs).
   • The LLM extracts requirement statements, identifies key terms, and structures them into PLM objects.
2. Initial Architecture Synthesis:
   • The LLM analyzes the extracted requirements collectively to infer system components (nouns) and interfaces (relationships).
   • It generates an initial Design Structure Matrix (DSM), where rows/columns represent components and cells represent interactions.
3. Human-in-the-Loop Refinement:
   • Engineers review, edit, and validate the AI-generated requirements and DSM.
   • This step ensures domain knowledge is correctly captured and corrects AI hallucinations or omissions.
4. Linking with CAD:
   • As designers create detailed 3D CAD models, the system establishes a one-to-one correspondence between the physical geometry and the abstract system elements in the MBSE model.
   • This links semantic requirements directly to geometric features (e.g., a specific flange geometry linked to an "attachment" requirement).

Verification and Validation (V&V)

The framework employs a hierarchical V&V approach:

• Phase 1 (Cross-Modal Compatibility): A rule engine checks consistency across modalities (e.g., ensuring geometric alignment matches functional constraints and that connection types are valid based on system graphs).
• Phase 2 (Formal Verification): Natural language constraints are translated into formal specifications using symbolic AI. System models are then mathematically verified against these specifications to ensure behavioral correctness.

3. Key Contributions

1. Unified Multimodal Data Model: A novel architecture that links semantic (text), geometric (B-rep), and relational (graph) data via UIDs and function/behavior-based mappings, moving beyond simple file association.
2. AI-Assisted Synthesis Workflow: A demonstrated methodology for using LLMs and Vision-Language Models (VLMs) to automatically generate initial system architectures (DSMs) from unstructured requirements, significantly reducing manual decomposition time.
3. Hierarchical V&V Framework: A dual-phase verification system combining cross-modal rule checking with formal symbolic methods to ensure system integrity.
4. Open-Source Implementation Strategy: The framework is designed around open-source geometry kernels and PLM capabilities, promoting reproducibility and accessibility for the scientific community.

4. Anticipated Results and Applications

While the paper presents a conceptual framework, it outlines anticipated impacts and illustrative applications:

• Realizing the Digital Thread: Enables end-to-end traceability where an engineer can click a CAD feature to see its constraints or query the impact of a requirement change across the entire system.
• Accelerated Architecture Exploration: Allows system architects to generate and iterate on multiple candidate architectures in hours rather than weeks, facilitating broader design space exploration in early project phases.
• Case Study (CubeSat Design): The framework is illustrated using a CubeSat scenario (Figure 3), demonstrating how it can manage tight mass/power budgets and complex interface requirements (e.g., launch vehicle constraints) by continuously checking subsystem compliance.

5. Significance

The GenAI Workbench represents a shift toward "MBSE 2.0"—an integrated, intelligent paradigm for systems engineering. Its significance lies in:

• Bridging the Semantic Gap: It solves the long-standing challenge of connecting unstructured natural language requirements with precise geometric and relational data.
• Enhancing Engineering Effectiveness: By automating the tedious tasks of requirement extraction and initial architecture generation, it allows engineers to focus on high-value validation and optimization.
• Scientific Validation: It serves as a scientific instrument to prove that multimodal data interoperability and generative synthesis workflows are computationally feasible, paving the way for future autonomous engineering agents.

Future Work involves completing the full implementation, conducting formal user studies to quantify benefits, fine-tuning VLMs on engineering datasets to deepen the physical-semantic bridge, and moving toward fully autonomous agentic verification.