Performance-based 3D object generation for engineering under limited data

DOI
URI
Scopus ID
Conference
IEEE Conference on Artificial Intelligence (CAI 2026) ; Granada, Spain ; 08–10 May 2026
Project
Publisher
IEEE
Book title
2026 IEEE Conference on Artificial Intelligence (CAI)
ISBN
First page
654
Last page
659
Peer-reviewed
✅
Part of the university bibliography
✅
Language
English
Keyword
Modeling
Printing
Training
Head
Bridges
Decoding
Three-dimensional displays
Clouds
Conferences
Limiting
Abstract
Designing 3D objects often requires meeting specific physical performance constraints. These objects are traditionally designed through iterative, resource-intensive processes. Generative models offer promising alternatives for generating new designs conditioned on physical conditions. However, generative design remains challenging due to the scarcity of 3D datasets labeled with continuous numerical conditions and the tendency of generative models to generate results that deviate from the target conditions, especially in data-limited scenarios. In this work, we propose a conditional variational autoencoder (CVAE) with a twin-head architecture to learn from small target datasets, we refer to it as the Twin-Head-CVAE. Our approach captures physical relationships in a shared latent space by using a source dataset with a smaller target dataset. We evaluate our approach on 3D point cloud representations of artificial concrete bridge structures using 1%, 10%, and 80% of the dataset as training set. Our results demonstrate a promising approach for generative design in scenarios with limited training data, which also enables the generation of 3D objects beyond the conditions present in the initial training data.
Version
Published version
Access right on openHSU
Metadata only access