FllumaOne: A Code-Native Multimodal CAD Dataset with Executable Programs and Kernel-Validated Feature Histories
FllumaOne introduces a multimodal CAD dataset with executable Python programs validated via OpenCASCADE, enabling high-precision geometric and feature-based learning.
Key Findings
Methodology
This work proposes a comprehensive CAD dataset built on executable Python programs within the Flluma environment, based on OpenCASCADE. Each sample pairs a structured feature tree, a replayable Python script, STEP geometry, surface point clouds, and natural language descriptions. Samples are accepted only after passing geometric validation, export success, and integrity checks, ensuring high-quality data. The dataset employs a controlled template library to generate models across four complexity levels, balancing diversity and manageability. A baseline model, Qwen2.5-Coder-1.5B LoRA, trained on 80,000 samples, achieves 99.98% Python syntax validity, 99.97% build success, and 99.14% STEP export validity on a 10,000-sample test set. Surface predictions yield a mean normalized Chamfer distance of 0.002124, demonstrating accurate geometric reconstruction.
Key Results
- On the test set, the baseline model achieved 99.97% geometric success and 99.14% STEP export success, with surface point cloud predictions having a mean normalized Chamfer distance of 0.002124, indicating near-perfect surface reconstruction.
- Multi-modal validation shows that combining geometric, visual, and textual modalities significantly improves feature tree prediction and conditional reconstruction accuracy.
- The dataset covers a broad complexity spectrum, from primitive shapes to complex mechanical parts, validating its utility across diverse CAD tasks.
Significance
FllumaOne bridges the gap between static geometry datasets and editable, construction-aware CAD models. Its multimodal, validated approach supports advanced tasks like program synthesis, feature prediction, and reverse engineering, fostering progress in AI-driven CAD design. By ensuring geometric correctness and providing rich multi-representational data, it enhances model interpretability, robustness, and industry relevance. This work addresses longstanding challenges in CAD data availability, paving the way for more intelligent, automated design workflows and manufacturing innovations.
Technical Contribution
The paper introduces a novel framework combining structured feature trees, executable Python scripts, and OpenCASCADE-based geometric validation, ensuring each sample's correctness. It employs a multi-layered complexity control strategy, balancing model diversity and data quality. The integration of multi-modal data—geometry, point clouds, images, and text—enables comprehensive learning and retrieval tasks. The baseline demonstrates high validity rates, validating the dataset's quality and potential for multi-task learning, setting a new standard for CAD data resources.
Novelty
This is the first dataset to unify executable Python programs, structured feature trees, and kernel-validated geometry within a large-scale, multimodal CAD resource. Unlike prior datasets focusing solely on geometry or images, FllumaOne ensures multi-representational alignment, enabling diverse tasks like program synthesis, reverse engineering, and multimodal retrieval. Its validation pipeline guarantees geometric correctness, making it uniquely suitable for industrial applications and research in editable CAD modeling.
Limitations
- The dataset relies on a predefined template library, which may limit the diversity of highly innovative or unconventional designs.
- Geometric validation, while rigorous, may still encounter failures with extremely complex topologies or non-standard structures, affecting sample diversity.
- Current focus is on static geometry and parametric models; future integration of material properties and dynamic behaviors is needed for broader industrial applications.
Future Work
Future directions include expanding the template library to cover more complex mechanical and freeform designs, integrating physical simulation and material data, and developing interactive editing tools. Enhancing the dataset with dynamic and deformable models, as well as real-world scan data, will further bridge the gap between synthetic and real industrial scenarios. Additionally, exploring unsupervised or semi-supervised learning approaches using this rich multimodal data could unlock new capabilities in AI-driven CAD automation.
AI Executive Summary
FllumaOne represents a significant advancement in CAD data resources by integrating multi-modal information with executable, validated programs. Traditional CAD datasets primarily focus on final geometries, lacking the detailed construction history and multi-representational data necessary for training intelligent models capable of editing and reverse engineering. This work addresses these limitations by constructing a large-scale dataset of 100,000 samples, each paired with a structured feature tree, a replayable Python script, STEP geometry, surface point clouds, and natural language descriptions. The models are generated through a controlled pipeline that employs a rich template library, ensuring diversity while maintaining geometric validity via OpenCASCADE validation. The baseline model trained on this dataset achieved near-perfect success rates, with 99.98% Python syntax validity and 99.14% STEP export success, demonstrating the dataset's high quality and practical utility. The surface reconstruction results, with a mean Chamfer distance of 0.002124, further validate the geometric fidelity of the data. This dataset opens new avenues for research in CAD program synthesis, feature prediction, multimodal retrieval, and reverse engineering, supporting tasks from simple parametric models to complex mechanical assemblies. Its multimodal nature allows models to leverage geometric, visual, and textual cues simultaneously, fostering more robust and interpretable AI systems. Looking ahead, expanding template diversity, incorporating material and physical properties, and enabling interactive editing are promising directions. Overall, FllumaOne sets a new standard for editable CAD data, promising to accelerate AI-driven automation and innovation in industrial design and manufacturing.
Deep Analysis
Background
近年来，深度学习在几何理解和生成方面取得了显著进展，推动了点云、网格和边界表示等数据集的发展（如ShapeNet、ModelNet、Thingi10K）。然而，这些数据多为静态几何模型，缺乏建模过程、参数依赖和操作历史，限制了模型的可编辑性和泛化能力。ABC、UV-Net等引入边界表示（B-Rep）和操作序列，提升了几何精度，但仍未充分结合多模态信息和几何验证机制。工业界对可验证、结构化的CAD模型需求日益增长，特别是在逆向工程、程序合成和自动化设计方面。近年来，融合语言、图像和点云的多模态数据集逐渐出现，但大多缺乏可执行性和几何验证，难以满足工业级应用的严苛要求。本文提出的FllumaOne结合了结构化特征树、Python可执行程序和OpenCASCADE几何核验证，为工业级可编辑CAD模型的学习提供了新平台。
Core Problem
现有CAD数据集多集中于静态几何或有限操作序列，缺乏多模态、多任务支持，难以实现复杂建模、逆向工程和程序合成的目标。缺少几何验证机制导致模型生成的模型可能存在拓扑或几何错误，影响实际应用。工业界需要一个高质量、结构化、可验证且多模态的CAD数据集，以支撑智能设计、自动化制造和逆向工程的需求。如何保证样本的几何有效性、操作的可重现性和多模态信息的同步，成为亟待解决的核心问题。
Innovation
本研究的创新点主要包括：1）引入结构化特征树和可执行Python程序，提供多层次的建模表示，兼顾可读性和可重现性；2）采用OpenCASCADE几何核验证，确保每个样本的几何有效性和拓扑正确性；3）设计多模态验证机制，将几何、点云、图像和文本信息同步，支持多任务学习；4）通过复杂度层级控制，丰富模型的表达能力，适应不同复杂度的建模场景。这些创新实现了高质量、多模态、多任务的CAD数据集，为深度学习在工业设计中的应用提供了坚实基础。
Methodology
- �� 样本生成：利用预定义的模板库，随机参数化生成多样化模型；• 结构化特征提取：构建有序的特征树，记录操作、参数和依赖关系；• 程序执行：在Flluma环境中运行Python脚本，重建几何模型；• 几何验证：使用OpenCASCADE核验证几何拓扑的有效性，包括面、边、点的合法性；• 多模态导出：生成STEP几何、表面点云、八视角图像和文本描述；• 样本筛选：仅采纳验证通过的样本，确保数据质量；• 归档：存储程序、特征树、几何、点云和多模态信息，形成完整样本集。
Experiments
在80,000个样本上训练基线模型，测试集包含10,000个样本，评估指标包括Python语法有效率、几何成功率、STEP导出成功率和点云重建误差。通过不同复杂度层级验证模型的泛化能力，进行消融实验分析多模态信息对重建精度的贡献。模型参数调优包括学习率、批次大小和训练轮次，确保结果的稳健性。模型在多任务场景中的表现验证了数据集的实用性和多样性。
Results
模型在测试集实现了99.97%的几何成功率和99.14%的STEP导出成功率，点云预测的平均归一化Chamfer距离为0.002124，显示出极高的几何重建精度。多模态验证表明，结合几何、视觉和文本信息能显著提升特征树预测和条件重建的准确性。不同复杂度层级的样本均表现优异，验证了数据集的多任务适应性和模型的泛化能力。
Applications
该数据集支持条件重建、程序合成、特征预测、逆向工程和多模态检索等多项工业应用。可用于自动化设计、个性化定制、智能制造和工业仿真，为工业设计提供高效、可解释的深度学习基础。结合物理仿真和材料属性，未来有望实现更复杂的工业场景应用。
Limitations & Outlook
目前依赖模板库，可能限制模型在极端或创新设计中的表现。几何验证在复杂拓扑模型中仍存在失败风险，影响样本多样性。模型主要针对静态几何，未来需考虑动态变形和材料特性，提升实际应用的适应性。
Plain Language Accessible to non-experts
想象你在一个工厂里做模型设计。以前，设计师用手工画图，最后得到一个模型，但这个模型不能告诉你怎么一步步做出来的，也不能修改。现在，FllumaOne就像给这个工厂配备了智能机器人，它不仅能画出模型，还能告诉你每一步怎么做，甚至可以用简单的指令让机器人帮你重新设计。它用一种特殊的语言写程序，机器人会按照程序一步步建出模型，还能验证每一步是否正确。这样，你可以随时修改设计，机器人也能理解你的意图。这就像有了一个会说话、会思考的设计助手，让复杂的模型变得像拼积木一样简单、可控。
ELI14 Explained like you're 14
想象你在玩一款建造游戏，你可以用不同的积木搭出各种东西。以前，游戏只告诉你最后的样子，但你不知道怎么一步步搭建的。现在，这个新系统就像给你一个神奇的说明书，不仅写出了每一步怎么搭，还能让你用简单的指令告诉它你想要的东西。它会用一种特殊的语言写出建造的步骤，然后用电脑把这些步骤变成真实的模型。更厉害的是，它还能检查每一步是不是做得对，确保模型没有错误。这样，你可以随意修改设计，系统也能理解你的想法，变得更聪明、更好玩。
Abstract
Parametric computer-aided design records both final geometry and the ordered construction history that determines how a part can be edited. Datasets for editable CAD research should therefore expose modeling operations, parameters, and feature dependencies together with validated geometry. We introduce FllumaOne, a code-native multimodal CAD dataset whose models are generated by executable Python programs in Flluma, a Qt/C++ OpenCASCADE-based CAD system. Each sample aligns its program with a structured feature tree, a training-oriented intermediate representation, STEP geometry, a surface point cloud, natural-language descriptions, metadata, and eight canonical visible-edge renderings. The primary release, FllumaOne-100K, contains 100,000 accepted samples across four template-level complexity regimes. Programs are executed and retained only after kernel geometry, solid validity, and export checks; release reports also record modality completeness and split-level duplicate tests. A Qwen2.5-Coder-1.5B LoRA baseline trained on 80,000 samples achieves 99.98% Python syntax validity, 99.97% Flluma build success, and 99.14% STEP-export validity on the held-out 10,000-sample test split. For the 9,909 predictions converted to surface point clouds, the mean normalized Chamfer Distance is 0.002124. The dataset supports conditioned CAD reconstruction, executable program synthesis, feature-tree prediction, B-Rep analysis, retrieval, design completion, and editable reverse engineering.