Neural Generation of Parametric CAD from 3D Geometry
under Data Scarcity Through Synthetic Data Generation
Name
yu-yunomi-sm-meche-2026-thesis.pdf
Size
13.22 MB
Format
Adobe PDF
Checksum (MD5)
64e51cc8c5d0815f13d5bf76756cd039
Author(s)
Yu, Nomi
Advisor(s)
Ahmed, Faez
Date Issued
February 2026
Publisher
Massachusetts Institute of Technology
Abstract
Mechanical design is complex and tedious, even for experts. AI has shown a strong potential for accelerating the design process with its recent successes in rapidly generating text, image, and even 3D content. However, it still struggles in generating practical designs in the engineering domain, particularly with respect to creating precise and editable outputs. CAD programs, structured as parametric sequences of commands that compile into precise 3D geometries, are a valuable target for automated generation given their precision, editability, and ubiquity in engineering. But generating these programs, especially from nonparametric data such as pointclouds and meshes, remains a crucial yet challenging task, typically requiring extensive manual intervention. Current deep generative models aimed at automating CAD generation are significantly limited by imbalanced and insufficiently large datasets, particularly those lacking representation for complex CAD programs. While private companies do hold large datasets of CAD models, they are not incentivized to share them to protect confidential data; and even if they were, the heterogeneity between different company datasets makes it difficult to train a coherent model over all data. As a result, generative CAD models still lack the ability to create complex, "realistic" models. We propose two frameworks utilizing synthetic data generation to address this data scarcity. First, we introduce GenCAD-3D, a multimodal generative framework utilizing contrastive learning for aligning latent embeddings between CAD and geometric encoders, combined with latent diffusion models for CAD sequence generation and retrieval. Additionally, we present SynthBal, a synthetic data augmentation strategy specifically designed to balance and expand datasets, notably enhancing representation of complex CAD geometries. Our experiments show that SynthBal significantly boosts reconstruction accuracy, reduces the generation of invalid CAD models, and markedly improves performance on high-complexity geometries, surpassing existing benchmarks. These advancements hold substantial implications for streamlining reverse engineering and enhancing automation in engineering design. Second, we introduce FedLDM, a federated learning framework for training conditional latent diffusion models tailored to engineering applications. FedLDM supports multi-stage generative architectures and uses synthetic data generation to improve robustness across private and heterogeneous datasets. We apply our method to the task of conditional CAD program generation from point clouds, using a two-stage architecture consisting of a CAD autoencoder and a diffusion transformer trained via classifier free guidance, enabling both synthetic generation and conditional inference. Our experiments demonstrate that FedLDM improves accuracy and achieves reconstruction accuracy within 18% of the "ideal" baseline. In particular, our method of utilizing synthetic generation achieves up to 29% improvement in reconstruction accuracy and over 56% improvement in valid generation rates relative to federated baselines. These results demonstrate an effective pathway for scalable, privacy-preserving generative design in engineering domains. Overall, we showed that our models are successful at parametric design generation from 3D inputs, and we demonstrated how synthetic data generation is an effective tool to improve the accuracy and validity of generative engineering models even in data-scarce situations. Our methods are a strong step in enabling engineers to efficiently and comprehensively explore the design space and accelerating the design cycle.
MIT Department
Massachusetts Institute of Technology. Department of Mechanical Engineering
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
