.png){kind=icon}
.png){kind=icon}
Return
Dynamic feature fusion for lightweight material recommendation in CAD assemblies
.png){kind=icon}
PRE
AI
Z
J
W
H
B
DOI：10.1007/s13042-026-03258-3.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Material selection in computer-aided design assemblies affects product performance, manufacturability, cost, and sustainability. Existing graph-based material recommendation methods can use assembly relations, but their repeated message passing increases computational cost and may blur component-level cues when geometrically similar parts have different material labels. This paper presents a lightweight feature-embedding framework for node-level material prediction in CAD assemblies. The model projects semantic, geometric, and physical component descriptors into an expanded embedding space, concatenates the learned embedding with the original descriptors, and uses a zero-initialized adaptive residual branch to control low-level feature supplementation during training. The design uses established projection, concatenation, residual, and gating operations, and its contribution is their task-specific integration into a compact CAD material recommendation pipeline. Experiments on the Fusion 360 CAD assembly dataset show improved Micro-F1 under multiple material-tier settings with computational cost close to an MLP-based predictor.
Keywords:
Material recommendation
Lightweight feature embedding
Feature fusion
CAD assemblies
Material classification
Journal
IF:
2.7
Papers:
3.1K
Citations:
5.6K
Organization
Cited Papers