Title: KDH-CAD: Knowledge-data hybrid CAD learning under data scarcity

Abstract:

The application of deep learning to computer-aided design (CAD) is primarily hindered by a significant lack of data. Obtaining large-scale, authentic CAD datasets is notoriously difficult, whereas synthetic alternatives often fail to accurately mirror genuine design workflows. Instead of attempting to amass ever-growing CAD datasets, this study reframes CAD learning as a problem of knowledge completion and calibration. To this end, we propose KDH-CAD, a hybrid framework that leverages a minimal set of labeled CAD data alongside structured domain knowledge sourced from textbooks and tutorials, integrated with pretrained knowledge from foundation models.

In this approach, domain knowledge serves to elicit and complete CAD-relevant concepts that are either under-represented or weakly expressed within pretrained foundation models. Subsequently, the scarce labeled CAD data is utilized to calibrate these concepts within the latent space, thereby accounting for task-specific geometric variability. Crucially, this calibration occurs without the need to fine-tune the foundation model itself.

Empirical evaluations on real-world mechanical part classification tasks demonstrate that KDH-CAD delivers robust performance even in low-data scenarios. The model achieves an accuracy of 92.6% using only 250 training samples, rising to 95.8% with 1,000 samples, and continues to benefit from additional data. These results are comparable to or surpass state-of-the-art methods, which typically demand an order of magnitude more training data. Consequently, these findings indicate that merging pretrained foundation models with structured domain knowledge can significantly diminish the dependency on large-scale CAD datasets, offering a principled and practical pathway toward data-efficient CAD learning.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC