Convolutional Autoencoder-based dimensionality reduction in Stratified Turbulence

Abstract:

Stratified turbulence is present in many nonlinear dynamical systems on Earth such as in the atmosphere and oceans. These flows are characterised by the presence of wide range of spatial and temporal scales, along with large scale instabilities or drafts. High fidelity numerical techniques such as Direct Numerical Simulations (DNS) are paramount to characterisation of such flows. However, such high resolution three-dimensional solvers produce large amount of data. For example, a simulation on a 512³ grid with single precision requires more than 0.5Gb of storage per variable, per time-step. To tackle this problem, dimensionality reduction can be used to retain a high temporal cadence in the data.

In order to compress and reconstruct such highly nonlinear data, we employ convolutional autoencoders (CAE). They are unsupervised feed-forward neural networks that aim to obtain an accurate approximation of the input data through a data compression and recovery process. In this work, we apply the CAE network to learn a DNS dataset of a stratified flow case exhibiting large scale instabilities. With an accurate lower-order approximation, we are able to store data with a high temporal resolution.