Masked Topology Modeling for Self-Supervised Learning on Parametric CAD
Proposes Masked Topology Modeling (MTM) for self-supervised learning on CAD B-rep graphs, achieving state-of-the-art results in geometric and topological understanding.
Key Findings
Methodology
MTM employs face-adjacency graphs, masking a subset of edges, and uses message passing to extract face features. A lightweight MLP predicts the masked edges' convexity and curve type, with labels automatically generated by the geometry kernel. Combined with MoCo contrastive learning and region reconstruction, the approach leverages multiple data sources, including the ABC dataset and synthetic CAD data, to pretrain a 27.9M-parameter encoder based on UV-grid, GATv2, and Transformer architectures.
Key Results
- On benchmarks like F360, SolidLetters, and MFInstSeg, MTM achieves top performance, with a macro mIoU of 84.55% on F360, surpassing previous state-of-the-art methods. In few-shot settings, the model maintains high accuracy, e.g., 68.92% mIoU with only 8 support samples, demonstrating excellent data efficiency. Ablation studies confirm that edge prediction enhances topological discrimination, especially in complex geometries.
Significance
This work addresses the scarcity of large-scale labeled CAD data by leveraging intrinsic geometric and topological cues for self-supervised learning. It enables models to understand complex structures without manual annotations, facilitating applications in automated design, reverse engineering, and defect detection, thus advancing industrial AI and smart manufacturing.
Technical Contribution
Introduces a novel edge masking prediction task on face-adjacency graphs, utilizing geometric kernels for label generation, and integrates contrastive learning with regional reconstruction. Theoretical analysis proves the identifiability of convexity and the partial non-identifiability of curve types, highlighting the model's ability to capture discriminative topological features. The architecture combines UV-grid, GATv2, and Transformer modules for robust feature extraction, supporting large-scale pretraining.
Novelty
First to formulate graph edge prediction as a self-supervised task in CAD B-reps, leveraging automatically generated labels from geometric kernels. Unlike prior face or node feature reconstruction, this approach emphasizes topological edge relations, significantly improving the model's capacity to understand complex structures and topological variations, surpassing existing face-level methods.
Limitations
- The approach's robustness diminishes with highly complex or non-manifold topologies, especially under severe geometric deformations. Pretraining relies heavily on synthetic data, and transferability to diverse industrial scenarios requires further adaptation. The edge classification is limited to convexity and curve type; extending to richer topological attributes remains future work. Computational costs are non-trivial, necessitating optimization for real-time applications.
Future Work
Future directions include multi-scale topology modeling, integrating physical constraints, and developing interactive CAD tools with real-time feedback. Extending the framework to handle more complex topological relations and non-manifold geometries, as well as improving computational efficiency, will broaden industrial applicability.
AI Executive Summary
This paper introduces Masked Topology Modeling (MTM), a novel self-supervised learning approach tailored for boundary representation (B-rep) CAD models. In industrial design, CAD data is often proprietary and scarce, limiting the application of data-hungry AI methods. MTM addresses this by exploiting the intrinsic geometric and topological structures of B-reps, specifically the face-adjacency graph, to generate rich supervisory signals without manual labeling.
The core idea involves randomly masking a subset of edges in the face-adjacency graph, then training a model to predict the hidden edges' convexity and curve type solely from the remaining face features. This is achieved through message passing over the graph, followed by lightweight MLP heads that classify edge attributes. The labels are automatically derived from the geometry kernel, ensuring no manual annotation is needed. To enhance feature discrimination, MTM is combined with MoCo contrastive learning and a region-based masked reconstruction task, which encourages the model to infer local geometric details.
Pretraining on large datasets, including the public ABC CAD dataset and a synthetic procedural dataset, demonstrates significant improvements over existing methods. Empirical results show that MTM achieves state-of-the-art performance on multiple benchmarks, such as F360 and SolidLetters, with notable gains in low-label regimes. Theoretical analysis confirms the model's ability to distinguish complex topological features, especially convexity, and highlights its superior discriminative power compared to face-level approaches.
This work marks a substantial step forward in CAD understanding, enabling models to learn from the structure itself rather than relying on extensive annotations. Its implications span automated design, reverse engineering, and quality inspection, fostering more intelligent and autonomous CAD workflows. Future research will focus on extending the framework to handle more complex topologies and integrating physical constraints, aiming for broader industrial deployment.
Deep Analysis
Background
随着工业数字化的发展，CAD成为设计和制造的核心工具。基于边界表示（B-rep）的参数化模型因其精确的几何和拓扑描述而广泛应用。早期如UV-Net和BRepNet通过深度学习提升了几何特征表达，但多依赖标注数据，限制了规模。近年来，自监督学习逐渐兴起，尝试利用几何结构自动生成训练信号，但在复杂拓扑关系捕获方面仍有限。CAD数据的私有性和稀缺性成为制约因素，促使研究者探索低标注依赖的学习策略。本研究基于面邻接图，提出利用边掩盖预测作为自监督信号，弥补了拓扑理解的不足。
Core Problem
工业CAD模型中，复杂的几何和拓扑关系难以用传统方法充分捕获，尤其在缺乏大规模标注数据时。现有面级或节点级学习方法无法有效理解边界的拓扑关系，导致在复杂结构识别和重建任务中表现有限。如何利用几何核自动生成标签，设计有效的自监督目标，成为关键挑战。同时，模型需在保持几何细节的同时理解拓扑结构，提升泛化能力和判别能力。
Innovation
本研究的主要创新在于引入面邻接图的边掩盖预测任务，将拓扑边作为预测目标，利用几何核自动生成标签，实现无标注的自监督学习。结合对比学习和区域重建，增强模型对几何和拓扑的理解。理论上证明凸性预测具有可识别性，拓扑关系的判别能力优于面级方法。架构采用UV-grid、GATv2和Transformer，支持大规模预训练，显著优于传统监督和节点重建方法。
Methodology
- �� 构建面邻接图，将每个面作为节点，边代表共享边界。• 在训练中随机掩盖部分边，删除对应边信息，模型仅利用剩余面特征进行消息传递。• 通过MLP头预测掩盖边的凸性和曲线类型，利用几何核自动生成标签，无需人工标注。• 结合MoCo对比学习，利用正负样本增强特征判别能力。• 采用区域重建任务，掩盖连续面区域，重建几何和类别信息。• 预训练使用ABC和合成数据集，增强模型泛化能力。• 结合多尺度拓扑信息，提升复杂结构理解。
Experiments
在公开的F360、SolidLetters、MFInstSeg等数据集上进行评估，采用宏平均mIoU、面级分类准确率等指标。对比多种监督和自监督方法，进行消融分析验证边掩盖预测的有效性。预训练模型在少样本和低标签场景中表现优异，显著优于传统方法。多数据源预训练策略增强了模型的泛化能力，实验证明其在工业场景中的适用性。
Results
模型在F360任务中实现84.55%的宏平均mIoU，超越现有SOTA。少样本实验中，预训练模型在8支持样本下仍达68.92%的mIoU，验证了其数据效率。消融实验显示，边预测目标的引入提升了拓扑关系的捕获能力，模型能区分复杂几何结构中的正反折叠。合成数据预训练后，模型在真实工业数据中表现出更强的泛化能力，验证了自监督策略的有效性。
Applications
该方法适用于工业CAD自动化设计、逆向工程、缺陷检测等场景。模型能理解复杂拓扑关系，辅助设计优化和结构分析。无需大量标注，适合私有数据环境，提升工业AI的智能水平。未来可结合交互式设计工具，实现端到端的CAD智能辅助，推动智能制造。
Limitations & Outlook
当前模型在极端复杂拓扑或非连续表面上鲁棒性不足，处理大规模几何变形仍有限。预训练依赖大量合成数据，迁移到不同工业场景需调优。边类别有限，未来应扩展至更丰富的拓扑关系和几何属性。模型计算成本较高，需优化推理效率。未来需增强模型的泛化能力和鲁棒性，适应更复杂的工业应用。
Plain Language Accessible to non-experts
想象你在一家工厂里，工人们需要组装各种复杂的机器。每个零件都像一个面，连接的边像零件之间的缝隙。工人们知道每个缝隙是凸的还是凹的，还知道缝的形状是直线还是弧线。现在，工厂想让机器人自己学会识别这些缝隙的类型，而不用告诉它们具体的标签。它们用一种方法，先把一些缝隙的特征隐藏起来，然后让机器人根据剩下的部分推断出被隐藏的缝隙到底是凸还是凹，是直线还是弧线。这样，机器人就能学会理解机器的结构，帮助设计和检测缺陷。这个方法让机器人变得更聪明，能在没有人帮忙的情况下理解复杂的机械结构，就像人类工程师一样。
ELI14 Explained like you're 14
想象你在玩拼图游戏，但有些拼块被藏起来了。你的任务是根据周围的拼块，猜出藏起来的拼块是什么样的。比如，有一块拼块连接两个不同的部分，你知道它是凸的还是凹的，或者是直线还是弧线。你会观察周围的拼块，利用它们的形状和连接关系，猜出隐藏的拼块。这就像在帮机器人学习理解复杂的机械结构，它通过观察剩下的部分，猜出被藏起来的连接方式。这样，机器人就能更好地理解各种复杂的机械设计，像人一样聪明地帮忙设计、检测缺陷或者改进结构。是不是很酷？它就像一个超级侦探，能从线索中推断出隐藏的秘密！
Glossary
Face-Adjacency Graph（面邻接图）
一种描述CAD模型中面与面之间连接关系的图结构，用节点代表面，边代表共享边界。
本文利用面邻接图作为自监督学习的基础结构。
Convexity（凸性）
描述边界的弯曲方向，分为凸、凹、平滑和刀锋四类，反映材料的内外折叠状态。
作为边的预测目标之一，几何核自动生成标签。
Curve Type（曲线类型）
描述边界曲线的几何形状，如直线、圆弧、B样条等。
用于预测边的几何特征，提升拓扑理解。
Masked Topology Modeling（MTM）
一种自监督任务，通过掩盖图中的边，训练模型预测被隐藏边的拓扑和几何属性。
核心创新点，增强模型对拓扑关系的捕获能力。
MoCo（Momentum Contrast）
一种对比学习框架，通过维护动量队列实现大规模负样本的高效利用。
用于增强特征判别能力。
Open Questions Unanswered questions from this research
- 1 如何扩展MTM以支持更复杂的拓扑关系和几何属性，仍是未来研究的重点。当前模型在极端复杂或非连续表面上鲁棒性不足，需结合物理信息和多尺度建模以提升性能。
Abstract
Computer aided design (CAD) is ubiquitous: virtually any modern object was designed using editable CAD tools. However, with the shortage of available CAD datasets in its native editable and parametric format, boundary representation (B-Rep), it is ever more important to develop data-efficient methods for this domain. We present a new self-supervised pretraining task, Masked Topology Modeling (MTM), that leverages the face-adjacency graph, an induced structure unique to B-reps that the encoder can be asked to reconstruct. MTM masks a fraction of edges and trains a small head to predict each masked edge's convexity and curve type from the encoder's post-message-passing face features. We combine MTM with a MoCo-style momentum-queue contrastive learning over B-rep-aware augmentations, a BFS-connected face-region masked-reconstruction objective, and pretraining on the ABC dataset and our new procedurally generated dataset to show strong performance on a number of benchmarks.