Req2CAD: bridging functional requirements and parametric CAD models to support conceptual 3D design

Paper Title

Req2CAD: bridging functional requirements and parametric CAD models to support conceptual 3D design

Publication Info

• Topic area: Early-stage conceptual CAD design using generative AI and functional decomposition.
• Keywords: Conceptual CAD, generative AI, function decomposition, parametric modeling, design ideation, human-AI collaboration, dual-feature representation, progressive CAD generation, design space exploration, function-structure reasoning.

Background and Problem

• Problem / challenge: Existing CAD systems have steep learning curves, constrain creativity through premature fixation, and lack effective support for transforming vague early-stage design concepts into parametric 3D models. Generative AI methods are limited in editability, reusability, and functional alignment of outputs.
• Significance: Addressing these challenges is critical for enabling novice designers to explore broader design spaces, align outputs with functional requirements, and improve the efficiency of conceptual 3D design workflows.
• Motivation and related work: Prior work includes generative design methods, AI-based CAD reconstruction, and text-to-CAD generation. However, these approaches often rely on domain-specific parameterizations, lack transparency, and fail to support iterative exploration or adequately align with functional requirements.

Solution

• Proposed approach: Req2CAD, an interactive system that integrates function decomposition, function–structure reasoning, and progressive CAD generation to assist novice designers in creating conceptual CAD models.
• Novelty:
  1. A data annotation pipeline linking functional requirements to 128k+ CAD components using vision-language models (VLMs) and large language models (LLMs).
  2. A dual-feature CAD representation combining boundary representation (B-rep) graphs and point cloud embeddings for geometric and topological exploration.
  3. An LLM-based progressive CAD generation method supporting multi-modal intent expression through text prompts and graphical interactions.
• Procedure and key techniques:
  1. Function decomposition: Designers input a design problem, and Req2CAD uses LLMs to decompose it into sub-functions in "verb + object" form.
  2. Function–structure reasoning: Req2CAD retrieves and clusters CAD components from a knowledge base based on functional alignment and similarity.
  3. Progressive CAD generation: Designers iteratively generate and refine CAD models using text prompts, graphical face selection, and incremental code generation in CADQuery.

Results

• Concrete findings:
  • Function annotation pipeline achieved 70%+ accuracy in mapping functions to CAD components, with higher efficiency than manual annotation.
  • Progressive CAD generation significantly reduced geometric errors compared to one-step generation, achieving 100% compile rate and improved alignment with design intent.
  • Designs created with Req2CAD scored higher in functionality (3.88 vs. 3.29) and novelty (3.31 vs. 2.75) compared to a baseline system.
• Advantage over baselines: Req2CAD outperformed the baseline (ChatGPT + Onshape) in supporting broader design exploration, intent expression, and transparency, while reducing frustration and improving satisfaction with design outcomes.
• Experiments / evaluation:
  • Technical evaluations: Function annotation pipeline tested on 50 CAD components; CAD generation methods compared on 120 samples.
  • User study: 12 novice designers completed two design tasks using Req2CAD and a baseline system, with expert evaluations of design quality and participant feedback on usability and task load.
• Limitations and future work:
  • Limited scope of the function–structure knowledge base; plans to expand with more complex components and assemblies.
  • Current focus on component-level modeling; future work will explore multi-component fusion and assembly.
  • Potential to incorporate additional inspiration sources (e.g., biomimicry databases) and support more modalities (e.g., gestures, voice).

Summary

Req2CAD is an interactive system that bridges functional requirements and parametric CAD models to support conceptual 3D design. By integrating function decomposition, a dual-feature CAD representation, and progressive CAD generation, it enables novice designers to explore broad design spaces and align outputs with functional needs. Technical evaluations and a user study demonstrated Req2CAD’s effectiveness in improving design quality, usability, and human–AI collaboration. Future work aims to expand the knowledge base, enhance multi-component assembly support, and explore additional modalities for intent expression.