An autoencoder-based bi-fidelity method utilising frozen latent spaces. Application to flow past a confined cylinder

This study presents a simple autoencoder-based bi-fidelity method, where the latent space is constrained to be invariant between both low- and high-fidelity datasets. The method is applied to the flow past a cylinder confined in a duct, where model fidelity is governed by computational model resolution. Significant cost savings are observed without significant compromise to approximation accuracy. The influence of several hyperparameters are considered including: latent space dimension, sample number and distribution. A method of generating synthetic solutions is also explored. This is done by performing regression on the common latent space of low- and high-fidelity models, which is then passed through the high-fidelity decoder. As a large number of samples are available from the low-fidelity model, significant improvements are observed compared with the traditional method of interpolating only high-fidelity data. • A bi-fidelity approach is proposed using autoencoders with a frozen latent space. • Present bi-fidelity strategy achieved better accuracy than latent-space regression. • An accurate generative model is obtained by interpolation in shared latent space.