Learning From Design Procedure To Generate CAD Programs for Data Augmentation
Proposes a design procedure-based prompting strategy using LLMs to generate industrial-grade CAD programs with high geometric diversity for data augmentation.
Key Findings
Methodology
This work introduces a multimodal prompting approach combining reference surface scripts and design process instructions to guide LLMs in generating complex CAD models. By defining B-spline reference surfaces via Python scripts and varying their parameters, the method enriches the geometric distribution of generated CAD programs. The prompts include system instructions, detailed design descriptions, and validation steps. Generated CAD programs are validated for structural validity and water-tightness, ensuring industrial applicability. Experiments utilize STEP format and B-spline ratio metrics, demonstrating significant improvements in geometric complexity and organic shape representation compared to baseline datasets.
Key Results
- Generated CAD models exhibit a B-spline face proportion of 77% and a curve proportion of 89%, with an average B-spline ratio of 0.2217, surpassing DeepCAD-b and GenCAD datasets. The models contain more complex organic geometries, with average line counts reaching 4494, face counts 26.57, and curve counts 67.57, indicating enhanced geometric richness.
- The diversity analysis shows a more uniform distribution over B-spline ratios, covering a wider range of shape complexities than baseline datasets, especially in higher B-spline regions.
- Ablation studies confirm that the reference surface prompts significantly influence the presence of B-spline geometries, validating the importance of the design procedure prompts for shape diversity.
Significance
This research addresses the critical bottleneck in industrial CAD data scarcity and geometric simplicity by enabling the synthesis of complex, organic-shaped CAD models. It bridges the gap between data-driven generative models and real-world industrial design standards, facilitating more realistic training data for deep learning models. The approach enhances the capability of LLMs to produce detailed, high-fidelity CAD programs, advancing automated design workflows and reducing manual effort. Its industrial relevance lies in improving design diversity, accelerating prototyping, and enabling smarter manufacturing pipelines.
Technical Contribution
The core innovation lies in integrating a script-based reference surface with a natural language-driven design process prompt, guiding LLMs to generate high-complexity B-spline geometries. The validation framework ensures the structural integrity of generated models. This approach differs from prior image-based or primitive-only methods, providing a parameterizable, scalable, and more industrially relevant data synthesis pipeline. It also demonstrates how multimodal prompts can effectively control complex geometric features in programmatic CAD generation.
Novelty
This is the first work to combine detailed reference surface scripts with design process prompts in LLM-based CAD program synthesis, significantly increasing geometric complexity and organic shape representation. Unlike previous methods relying solely on image prompts or primitive operations, this approach leverages script-based parametric guidance, enabling the generation of models with high B-spline content that closely resemble real industrial designs. It fills a critical gap in scalable, diverse CAD data synthesis for deep learning applications.
Limitations
- The method has been validated primarily on bracket-like components; extending to more complex assemblies or multi-part designs requires further research.
- Generation quality depends heavily on the pre-trained LLM's understanding of CAD scripting and design semantics, which may limit detail fidelity.
- Validation ensures water-tightness but does not directly address manufacturability constraints such as tolerances, material properties, or assembly considerations. Future work should incorporate physical simulation and manufacturing-aware optimization.
Future Work
Future directions include expanding to diverse categories such as automotive parts and consumer products, integrating multimodal inputs like images or point clouds for richer conditioning, and incorporating physical and manufacturing constraints into the generation process. Additionally, exploring reinforcement learning or self-supervised methods to improve the fidelity and diversity of generated CAD models will be valuable. These advancements aim to create a comprehensive, scalable pipeline for industrial CAD data synthesis and automated design.
AI Executive Summary
In modern industrial design, creating complex CAD models that meet both aesthetic and functional requirements remains a challenge. Existing datasets lack the geometric richness needed for training robust AI models, limiting automation potential. This study introduces a novel prompting strategy that leverages the design process, combining reference surface scripts and natural language instructions, to guide large language models in generating highly diverse, organic CAD programs. By defining reference surfaces through parametric B-spline scripts and varying their parameters, the approach significantly enhances geometric complexity, producing models with over 77% B-spline faces and 89% curves, outperforming existing open-source datasets.
The core innovation lies in integrating script-based reference surfaces with design process prompts, enabling the generation of models that closely resemble industrial-grade designs. Validation mechanisms ensure the structural integrity and water-tightness of the generated models, making them suitable for downstream manufacturing and simulation tasks. Experimental results demonstrate that the generated CAD models exhibit a broader distribution of shape complexity, with average line counts reaching 4494 and a more uniform B-spline ratio distribution, indicating higher geometric diversity.
This method addresses a key bottleneck in CAD data scarcity and complexity, providing a scalable, parameterizable pipeline for industrial data augmentation. Its implications extend to accelerating prototyping, reducing manual effort, and fostering the development of intelligent design tools. Future work aims to expand the approach to more complex assemblies, incorporate multimodal inputs, and embed manufacturing constraints, pushing the boundaries of automated industrial design and AI-driven CAD synthesis.
Deep Analysis
Background
随着工业设计对高复杂度CAD模型的需求不断增长，现有公开数据集如DeepCAD、GenCAD多偏向简单几何体，难以反映行业实际的几何多样性。传统方法依赖手工制作或有限操作集，难以实现规模化、多样化的工业级模型生成。近年来，深度学习和大语言模型（LLMs）在CAD程序自动生成中逐渐成为研究热点，但受限于数据单一和几何表达能力不足，难以满足工业应用的需求。引入工业设计中的参考曲面和流程提示，为自动化生成复杂有机形状提供了新思路，成为突破口。
Core Problem
现有CAD生成模型在几何复杂度和多样性方面存在明显不足，难以满足实际工业设计的需求。主要问题包括：缺乏多样化的几何表达，生成模型多为简单几何体；训练数据有限，难以捕捉行业中的复杂形状特征；生成程序缺乏对有机、自由曲面等复杂几何的控制能力。这些限制阻碍了自动化设计的推广和工业应用的普及。如何在保证模型多样性的同时，提升几何复杂度，是亟待解决的核心问题。
Innovation
本研究的创新点主要体现在：1）提出结合工业设计流程的多模态提示策略，将参考曲面脚本与设计步骤结合，指导LLMs生成高复杂度的B样条几何模型；2）利用Python脚本定义的参数化参考曲面，自动变换参数以丰富几何形态，提升模型多样性；3）引入结构和有效性验证机制，确保生成的CAD模型为水密实体，符合工业标准。这些创新突破了传统数据集的局限，显著提升了模型的表达能力和几何复杂度，为工业级CAD自动生成提供了新技术路径。
Methodology
- �� 设计流程提示：结合工业设计中的参考曲面（如高斯、鞍面、波浪面）和流程描述，构建多模态提示。
- �� 参考曲面定义：用Python脚本描述B样条曲面参数，自动变换参数以丰富几何形态。
- �� LLM生成：输入格式包括系统提示、设计描述、参考曲面脚本和验证指令，生成CAD程序。
- �� 结构验证：执行程序，检测水密性和拓扑合理性，确保实体质量。
- �� 反复优化：错误反馈调整提示，提升生成成功率。
Experiments
采用STEP格式的CAD模型作为输出，评估指标包括线数、面数、曲线数及B样条比例。对比DeepCAD、GenCAD等基线，验证几何复杂度和有机形状比例。通过不同参考曲面变换，分析模型多样性。实验中还进行消融分析，验证设计流程提示的作用。
Results
生成模型在几何复杂度方面优于公开数据集，平均线数达4494，面数26.57，曲线数67.57，B样条面占比77%，曲线89%，B样条比例0.2217。多样性分析显示，模型能生成丰富的有机形状，显著高于DeepCAD-b、GenCAD等。消融实验验证了参考曲面提示对提升几何复杂度的关键作用。
Applications
该方法适用于工业设计、机械制造、虚拟样机等场景，能自动生成复杂的工业级CAD模型，减少设计时间，提高设计多样性。结合物理仿真和制造约束，未来可实现端到端的自动化工业设计流程。
Limitations & Outlook
目前主要验证于支架类模型，泛化到更复杂类别仍需优化。模型依赖预训练LLMs能力，可能在细节表达上受限。验证机制虽确保水密性，但实际制造还需考虑材料和公差，未来需结合物理模拟优化。
Plain Language Accessible to non-experts
想象你在厨房里做菜，准备一道复杂的菜肴。你会先准备一些基本材料，比如蔬菜、肉类，然后根据食谱一步步加入调料、切割、烹饪。每一步都影响最终的味道和外观。现在，科学家们用类似的方法，先定义一个基础的“参考材料”——比如一个有机形状的面团，然后用计算机“食谱”指导它变成各种复杂的菜肴。这些“菜肴”就是复杂的3D模型。通过调整参考材料和步骤，模型可以变得更丰富、更接近真实工业设计的样子。这就像用不同的调料和烹饪方法，做出多样的菜肴一样。
ELI14 Explained like you're 14
想象你在玩乐高积木，但你想拼出超级酷的建筑。普通的积木只能拼出简单的房子，但你想要弯弯曲曲、像自然界一样的形状。科学家们用电脑帮忙，先定义一个基础的“模样”，比如一个弯弯曲曲的曲面，然后用特殊的“拼装指南”让电脑帮你拼出复杂的模型。这些指南就像是给电脑的“食谱”，告诉它怎么拼、用哪些特殊的积木。结果，电脑拼出来的模型比以前更复杂、更漂亮，就像用乐高拼出真正的艺术品一样。这个方法可以帮设计师更快、更好地创造出工业用的复杂零件，就像你用乐高拼出未来的城市一样酷！
Abstract
Large Language Models (LLMs) have demonstrated impressive capabilities in a wide range of code generation tasks. However, generating code for certain domains remains challenging. One such domain is Computer-Aided Design (CAD) program, where the goal is to produce scripted parametric models that define object geometry for precise design and manufacturing applications. A key challenge in LLM-based CAD program generation is the limited geometric complexity of generated shapes compared to those found in real-world industrial designs. This shortfall is in part due to the lack of diversity in the available CAD program training data. To address this, we propose a novel data augmentation paradigm that prompts an LLM to generate CAD programs conditioned on a reference surface program and a modeling procedure - an idea inspired by practices in industrial design. By varying the reference surface using a collection of organic shapes, our method enriches the geometric distribution of generated CAD models. In particular, it introduces edges and faces defined by spline-based curvature, which are typically missing or underrepresented in existing open-source CAD program datasets. Experiments show that our method produces CAD samples with significantly greater geometric diversity and a higher resemblance to industry-grade CAD designs in terms of the proportion of organic shape primitives. This enhancement makes our CAD data augmentation approach a useful tool for training LLMs and other deep learning models in CAD generation.