Spatial super-resolution of finite-time Lyapunov exponent fields with generative adversarial networks

Abstract

In this article, two generative adversarial networks that enhance the spatial resolution of two- and three-dimensional finite-time Lyapunov fields by a factor of 4 in each dimension are presented. Using these models, high-resolution distributions can be approximated based on low-resolution input, the latter of which is associated with a computational cost and storage reduced by factors of about 42 and 43, respectively. Evaluating various test cases, the performance and generalizability of this approach are assessed. Shortcomings are only observed in the case of high-frequency spatial fluctuations where no sufficient statistical information is available in the low-resolution space. The major flow structures, however, are adequately rendered, giving rise to Lagrangian analyses of complex flow configurations that may otherwise remain elusive due to an excessive computational cost.