Deep reliability learning with latent adaptation for design optimization under uncertainty

Deep learning may encounter the generalization issue for uncertainty propagation due to the limited number of training data, especially when the training and test sample points follow different probability distributions. To overcome this challenge, this paper presents a latent adaptation-based deep learning approach to enhance the accuracy of deep networks for reliability-based design optimization under data constraint. By collecting training data from a source domain, an autoencoder is first utilized to introduce a meaningful low-dimensional latent space. Then a feedforward neural network is trained to link the original inputs to latent variables. To address the generalization issue, a latent adaptation strategy is proposed to link the latent variables across the target and source domains. A Gaussian process model is then constructed to learn the limit state function for the target domain in the latent space. Monte Carlo simulation is employed for reliability analysis in target domains. The latent adaptation-based deep learning is then integrated with reliability-based design optimization, and the effectiveness of the proposed approach is demonstrated through two examples.(c) 2022 Elsevier B.V. All rights reserved.

Automated summary

This paper presents a latent adaptation-based deep learning approach to enhance the accuracy of deep networks under data constraint. By introducing latent space and a latent adaptation strategy, the generalization issue caused by different distributions of training and test samples is addressed.