A multi-modal CAD generation network for generating a CAD command sequence and a method of generating the same

By integrating text and geometric features through a multimodal CAD generation network and combining it with a group-relative strategy optimization algorithm, the problem of insufficient geometric accuracy and editability of existing CAD models in complex part design is solved, achieving CAD model generation with higher accuracy and reliability.

CN122174303APending Publication Date: 2026-06-09GUANGDONG UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
GUANGDONG UNIV OF TECH
Filing Date
2026-02-11
Publication Date
2026-06-09

AI Technical Summary

Technical Problem

Existing text-driven CAD model generation suffers from weak spatial geometry perception, unmanufacturable generated models, and lack of physical property feedback when dealing with complex industrial parts designs. This results in low geometric accuracy and editability of the generated CAD models, making them difficult to apply in practical engineering.

Method used

A multimodal CAD generation network is adopted, which combines a text encoder, a geometric feature extraction module, a spatially sparse cross-modal attention fusion module, and a two-stage CAD decoder. By fusing textual semantic features and geometric features, a CAD command sequence is generated. A group relative policy optimization algorithm is introduced for training to ensure that the generated CAD model conforms to physical and topological constraints.

Benefits of technology

It improves the geometric-physical accuracy and editability of generated CAD models, enabling the generated models to better conform to actual physical laws and assembly requirements, with a more robust topology, significantly improved success rate, and engineering application value.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122174303A_ABST
    Figure CN122174303A_ABST
Patent Text Reader

Abstract

The application provides a multi-modal CAD generation network for generating a CAD command sequence and a generation method thereof, and relates to the technical field of deep learning. A data set is constructed and divided into a training set, a validation set and a test set. Training is performed based on the training set, the generation effect is verified by using the validation set, an initially trained network is obtained, and the performance is tested by using the test set. Based on a group relative strategy optimization algorithm, the parameters of the initially trained network are adjusted to obtain a finally trained network. Based on a text description and an auxiliary 3D grid model, the input is processed by using a first text encoder module to extract text semantic features, a first geometric feature extraction module is used to extract geometric features, and a first spatial sparse cross-modal attention fusion module is used to fuse the text semantic features and the geometric features to obtain fused features. A first two-stage CAD decoder module is used to generate a CAD command sequence. The application improves the geometric-physical accuracy and editability of the generated CAD model.
Need to check novelty before this filing date? Find Prior Art