Towards High-Fidelity CAD Generation via LLM-Driven Program Generation and Text-Based B-Rep Primitive Grounding

1. Problem Definition

Current AI-driven Computer-Aided Design (CAD) generation methods suffer from a fundamental paradigm gap between two distinct approaches:

• Parametric Modeling: Generates CAD as a sequence of operations (e.g., extrude, revolve). While editable, it struggles with advanced features like fillets, chamfers, and shells because these operations require explicit selection of Boundary Representation (B-Rep) primitives (faces, edges) that do not exist in the parametric history until the operation is executed.
• Direct B-Rep Synthesis: Generates the final geometry and topology directly. While capable of complex shapes, it lacks the parametric structure and design intent, making models difficult to edit or reuse.

The Core Challenge: Modern feature-based CAD systems require a tight coupling where advanced operations (like a fillet) must explicitly select primitives from the transient B-Rep state generated by previous steps. Existing methods treat parametric generation and B-Rep selection as separate problems, leading to low fidelity, invalid models, and an inability to handle complex industrial designs.

2. Methodology: FutureCAD Framework

The authors propose FutureCAD, a unified framework that integrates Large Language Models (LLMs) for program generation with a specialized transformer module for grounding text queries to geometric primitives.

A. System Architecture

FutureCAD operates in a closed loop:

1. LLM Program Generation: An LLM generates an executable CadQuery script based on a natural language description.
2. Text-Based Querying: When a feature requires a primitive reference (e.g., "fillet these edges"), the LLM inserts a natural language query (e.g., cq.Query("All circular holes on the surface")) into the code instead of hard-coding indices.
3. B-Rep Grounding (BRepGround): During execution, the BRepGround module takes the current transient B-Rep state and the text query. It identifies and selects the specific geometric primitives (faces/edges) corresponding to the query.
4. Execution: The resolved primitives are fed back to the CAD kernel to execute the operation, updating the B-Rep state for the next step.

B. Key Components

• BRepGround (Transformer-based Grounding):

  • Text Encoder: Uses a pretrained BERT model to encode natural language queries.
  • B-Rep Encoder: Uses UV-Net to extract embeddings for faces (2D UV grids) and edges (1D UV grids), augmented with geometric attributes. A Graph Neural Network (GNN) propagates context across the face-adjacency graph.
  • Fusion Module: A transformer-based module using cross-attention to fuse text embeddings with B-Rep token embeddings. This allows the model to align semantic constraints (e.g., "top surface") with geometric topology.
  • Output: Predicts a score for each primitive to select the target set.

• LLM Training Strategy (Two-Stage):

  1. Supervised Fine-Tuning (SFT): Trains the LLM (Qwen2.5-7B) to generate CadQuery scripts from text descriptions. The dataset includes "rewritten" code to ensure diversity and natural language queries for primitive selection.
  2. Reinforcement Learning (RL): Uses Group Sequence Policy Optimization (GSPO). The reward function is based on the Chamfer Distance (CD) between the generated geometry and the ground truth. Failed executions receive a zero reward, penalizing invalid code.

C. Dataset Construction

The authors constructed a new dataset of 140,000+ real-world CAD models (from Onshape):

• Includes advanced features (fillet, chamfer, shell).
• Annotated with abstract and detailed geometric descriptions.
• Converted into executable CadQuery programs with text-based queries for primitive grounding.

3. Key Contributions

1. Unified Paradigm: Introduced FutureCAD, the first framework to unify LLM-driven parametric program generation with explicit text-based B-Rep primitive grounding, enabling realistic feature-based modeling.
2. BRepGround Module: Proposed a novel transformer architecture that effectively grounds natural language queries to specific geometric primitives within a dynamic B-Rep state, solving the "selection" bottleneck.
3. Large-Scale Dataset: Released a dataset of 140k+ complex CAD models with rich annotations for both text-to-CAD generation and primitive grounding tasks.
4. Training Strategy: Demonstrated that combining SFT with RL (using geometric rewards) significantly improves the validity and generalization of generated CAD programs.

4. Experimental Results

The method was evaluated on the FutureCAD dataset (Advanced and Standard subsets) and the out-of-distribution Fusion360 dataset.

• Text-to-CAD Performance:
  • Geometric Fidelity: Achieved the lowest Chamfer Distance (CD) (Median CD: 31.12 on Advanced, 14.68 on Standard), significantly outperforming baselines like CADFusion and CAD-LLaMA.
  • Validity: Achieved an extremely low Invalidity Ratio (IR) of 1.01% (Advanced) and 0.91% (Standard), compared to >40% for SFT-only baselines.
  • Generalization: Maintained high performance on the out-of-distribution Fusion360 dataset, proving robustness.
• B-Rep Grounding Performance:
  • Achieved 63.07% mAP and 50.23% F1 on the primitive selection task.
  • Outperformed baselines (including LLM-based code generation and CLIP-based approaches) by a large margin, demonstrating the necessity of explicit geometric-text alignment.
• Ablation Studies: Confirmed that removing BRepGround (replacing it with nearest-neighbor point selection) drastically increased the invalidity rate (to 8.29%) and reduced geometric fidelity, highlighting the importance of semantic grounding.

5. Significance

• Bridging the Gap: FutureCAD successfully bridges the gap between high-level design intent (natural language) and low-level geometric constraints (B-Rep primitives), a critical step for AI in industrial design.
• Practical Applicability: By generating valid, editable CadQuery scripts, the framework integrates seamlessly with existing professional CAD workflows, unlike direct B-Rep synthesis methods which produce "dead" geometry.
• Scalability: The approach demonstrates that LLMs, when augmented with specialized grounding modules and RL, can handle complex, multi-step design tasks requiring precise geometric reasoning, paving the way for accessible, AI-assisted engineering.