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

Technical Summary: KDH-CAD

Problem Statement
Deep learning in Computer-Aided Design (CAD) is fundamentally constrained by data scarcity. Unlike natural language or images, authentic industrial CAD data is difficult to collect at scale due to proprietary restrictions and the high expertise required for modeling. While synthetic datasets (e.g., ABC, SketchGraphs) exist, they often fail to capture real-world design intent, feature history, and functional constraints, limiting their applicability to practical scenarios. Consequently, learning-based CAD methods lag behind other data-rich fields. Existing approaches attempting to mitigate this via synthetic data scaling or few-shot learning often lack the ability to incorporate structured domain knowledge or suffer from the "concept-geometry mismatch," where geometric variability within a single part category confuses purely data-driven models.

Methodology: KDH-CAD
The authors propose KDH-CAD (Knowledge–Data Hybrid CAD learning), a framework that reframes CAD learning under data scarcity as a knowledge completion and calibration problem. Instead of relying on massive datasets, KDH-CAD integrates three components:

1. Pretrained Foundation Models: Specifically, a Vision-Language Model (VLM), Qwen3-VL-2B-Instruct, which provides a transferable latent representation space.
2. Structured Domain Knowledge: Sourced from mechanical design textbooks and tutorial videos, organized into "engineering knowledge triplets" (Canonical Terminology, Description, and Visual Demonstration).
3. Limited Labeled Data: A small set of task-specific CAD samples used for calibration.

The framework operates in two primary stages:

• Knowledge Eliciting and Completion:
  The VLM is conditioned on the structured engineering knowledge triplets to extract "knowledge embeddings." This process elicits CAD-relevant concepts (e.g., specific part types like gears or flanges) that are often weakly expressed or suppressed in the general pretraining corpus. By anchoring the VLM with authoritative definitions and visual examples, the model generates embeddings that represent canonical CAD concepts rather than generic visual features.

• Knowledge Calibration with Limited Data:
  To address the mismatch between abstract concepts and specific geometric variations in the target task, a lightweight Knowledge Calibration Network is introduced. This network keeps the foundation model frozen and learns two specific mechanisms:

  1. Shift-Vector Submodule: Learns additive shift vectors for each concept based on the target dataset's distribution. This re-centers the knowledge embeddings to account for geometric variability (e.g., different shapes of "brackets") without fine-tuning the massive foundation model.
  2. Cross-Attention Submodule: Aligns the shifted knowledge embeddings with the input data embeddings (derived from multi-view rendered images of CAD parts). The data acts as a query to weight and aggregate the most relevant knowledge concepts, producing a robust, task-aligned representation.

The system is instantiated for part classification, where it outputs both a category label and, uniquely, a semantic description of the part generated by the VLM based on the aligned knowledge.

Key Contributions

1. Framework Introduction: The proposal of KDH-CAD, a hybrid framework that combines pretrained foundation models, structured domain knowledge, and minimal labeled data to solve CAD data scarcity.
2. Novel Part Classification Method: A specific implementation achieving state-of-the-art performance with only hundreds of training samples, contrasting with existing methods that require tens of thousands.
3. Alternative to Data Scaling: An empirical demonstration that integrating structured domain knowledge is a practical and effective alternative to scaling up synthetic or real-world CAD datasets.

Experimental Results
The method was evaluated on two real-world mechanical part classification datasets: TMCAD (9,799 models, 10 categories) and FabWave (subset of 8 categories).

• Low-Data Performance: KDH-CAD achieved 92.6% accuracy with only 250 training samples and 95.8% with 1,000 samples on TMCAD. This performance matches or exceeds state-of-the-art baselines (UV-Net, AAGNet, BRT) that were trained on datasets 10 to 80 times larger.
• Robustness: The method demonstrated consistent accuracy across ten random data splits with negligible variance, indicating robustness to sample selection.
• Retrieval Capability: In CAD model retrieval tasks, KDH-CAD achieved a 96.18% Recall@1 with only 1,000 samples, outperforming baselines trained on 10,000 samples by nearly 10 percentage points.
• Semantic Output: Unlike baselines that only output class indices, KDH-CAD successfully generated interpretable semantic descriptions (category, geometry, function) for classified parts.
• Ablation Studies: Removing domain knowledge or the calibration modules (shift vectors and cross-attention) resulted in significant performance drops, confirming that both knowledge elicitation and geometric calibration are essential.

Significance and Claims
The paper claims that KDH-CAD provides a principled and practical direction for data-efficient CAD learning. By treating CAD learning as a knowledge completion task, the approach substantially reduces reliance on large-scale, high-quality CAD datasets, which are often difficult to obtain. The authors note that the method mirrors an engineering student's learning process: acquiring concepts from authoritative references and calibrating them through representative exercises.

However, the authors are modest regarding limitations. They acknowledge that the method may struggle with highly complex or uncommon parts because the generic domain knowledge (textbooks/videos) covers canonical structures but may miss specialized, application-dependent variants. They suggest that future work could incorporate proprietary, company-specific knowledge (potentially via federated learning) to address these edge cases. Additionally, the current implementation relies on rendered multi-view images rather than native CAD topology (e.g., B-rep), though the core framework is not inherently limited to image inputs.