Inverse structural design with generative and probabilistic autoencoders and diffusion models

Traditional structural design is a forward trial-and-error process. Designers need to iterate through different design solutions and conduct structural analysis until the design meets the codes and standards. This study proposes and investigates a generative machine learning (ML) framework for inver...

Full description

Saved in:
Bibliographic Details
Published in:Engineering applications of artificial intelligence Vol. 161; p. 112143
Main Authors: Zhuang, Bozhou, Gallet, Adrien, Smyl, Danny
Format: Journal Article
Language:English
Published: Elsevier Ltd 01.12.2025
Subjects:
Conditional autoencoders
Continuous beams
Denoising diffusion models
Generative artificial intelligence
Inverse problem
Structural design
Tabular data
Structural design
Generative artificial intelligence
Tabular data
Continuous beams
Conditional autoencoders
Denoising diffusion models
Inverse problem
ISSN:0952-1976
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Traditional structural design is a forward trial-and-error process. Designers need to iterate through different design solutions and conduct structural analysis until the design meets the codes and standards. This study proposes and investigates a generative machine learning (ML) framework for inverse design of continuous beam systems. Three generative ML models, including conditional variational autoencoder (CVAE), conditional autoencoder with maximum likelihood estimation (CAE-MLE), and denoising diffusion models (DDMs) are trained and fine-tuned on the CBeamXP (Continuous Beam Cross-section Predictors) dataset with 1,000,000 beam sections to generate the cross sectional properties. Research results show that CAE-MLE achieves the highest generation accuracy and robustness, while CVAE offers more variability through latent space sampling. DDMs provide controllable generation variability via a stochasticity parameter in the inverse diffusion process. The proposed framework enables efficient generation of multiple design solutions and can potentially accelerate the conceptual design workflows in structural engineering. This work also demonstrated the feasibility toward artificial intelligence (AI)-assisted structural design using generative approaches and tabular datasets. •A generative design framework is developed for continuous beam systems.•Three generative ML models, CVAE, CAE-MLE, and DDMs, are trained and fine-tuned.•The generative design performance of different models is compared.•The design variability is investigated under specific design conditions.
ISSN:0952-1976
DOI:10.1016/j.engappai.2025.112143