Autoencoders have recently gained popularity as a model order reduction tool in applications of fluid mechanics by compressing input high-dimensional data using an encoder to map the input space into a lower-dimensional latent space. In this study, we parameterize a forced isotropic turbulence flow into Statistical characteristics of turbulence by testing combinations of variational and neural encoder-decoder structures and performances are compared based on the reconstruction error for turbulent flows. The Data is extracted from the John Hopkins Turbulence database’s Forced Isotropic Turbulence (coarse) through their Matlab library. The extracted data is in the shape of 5000 64x64 x-y plane slices, taken from random points in the dataset where each cell contains 16 individual elements consisting of 3 channels for x, y, z velocities (u,v,w), 1 for pressure (p), 9 for velocity gradients (ux, uy, uz, vx, vy, vz, wx, wy, wz) and 3 for pressure gradients (Px, Py, Pz). These are selected in order to extract their dependencies on each other and help in the estimation and reconstruction of the flow after passing it through the Neural Networks. It is then fed to four different autoencoder models with different architectures that map this input data back to itself via a latent space. The latent space generated provides a way to control the flow through Statistical characteristics of turbulence. The latent vector of the autoencoder’s encoder end is supposed to be dependent upon the characteristics of the slice.&nbsp; These characteristics include the total kinetic energy, dissipation factor, root mean squared velocity, Taylor microscale factor, Taylor scale Reynolds number, Kolmogorov time scale, Kolmogorov length scale, Integral scale, and large Eddy turnover time, which is averaged over the spacial region given. The latent vector also contains independent entries which can be used to create different output with similar properties. After encoding, the networks can be broken to either extract the latent space parameters from the encoding end or generate a flow field dependent upon the input property vectors. The use of neural network-based architecture removes the need for dependency on the classical mesh-based Navier-Stoke equation estimation. If trained on more varied datasets and a large number of data slices, it can prove to be a model for creating and estimating turbulent flows. Potential application fields can range from professional CFD flow prediction and modeling in CAE application, which can be backed up by low-resolution traditional CFD analysis to creating video games / virtual environments for animation and other computer graphics applications.

Generation and Parameterization of Forced Isotropic Turbulent Flow Using Autoencoders and Generative Adversarial Networks

Paper Proceedings: https://asmedigitalcollection.asme.org/IMECE/IMECE2021/volume/85543

ASME 2021 International Mechanical Engineering Congress and Exposition (IMECE2021)

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.