Machine/Deep Learning

The increased computational effort inherent in DNS and other numerical formulations, both in time and hardware needs, has opened the way to adopt novel computational techniques, stemming from Machine Learning (ML). Aerodynamic coefficient prediction, turbulence modelling, transitional flow modelling, and flow reconstruction are just a few fields of application. To this direction, Deep Learning (DL) techniques have attracted a lot of interest, especially for their use as surrogate models in Computational Fluid Dynamics (CFD). DL models can deal with complex mathematical concepts through a layered network of interconnected nodes, and they can be easily applied on applications involving data of 2-D and 3-D dimensions, where the computational domain is discretized using meshes. In CFD, a dense grid leads to higher accuracy, and this corresponds to a deeper network in the DL method. Therefore, the alternative solution of employing DL methods can improve the flow estimation and diminish the computational time. A constantly evolving field of research lately with ever-increasing accuracy, Super Resolution (SR), utilizes high resolution (HR) data reconstruction using sparse measurements with DL approaches. To achieve this, convolutional neural networks (CNNs), Variational Autoencoders (VAEs) and generative adversarial networks (GANs) are among the most common architectures utilized. In the MOVEFREE research project, open channel flow image data derived from DNS simulations are employed to train a super resolution model, making it capable of upscaling sparse/coarse data onto a high-resolution (HR) field. The proposed architecture is based on consecutive CNN layers, in an encoding/decoding manner, with residual connections to ensure proper information exchange. The reconstruction results on our open channel flow DNS data are very close to the HR images, suggesting that the method can be employed in relevant fluid mechanics applications, as a complimentary method to time and resource-intensive DNS simulations.