Explore a comprehensive webinar on modelling and controlling turbulent flows through deep learning, led by Ricardo Vinuesa as part of the Data-Centric Engineering Webinar Series. Delve into advanced topics such as the effect of Reynolds number on well-resolved LES, adaptive simulations of NACA0012 profile, and high-fidelity simulations of wing-tip vortices. Discover various applications of machine learning in fluid mechanics, including flow reconstruction using convolutional neural networks (CNN), transfer learning techniques, and generative adversarial networks (GANs) for high-resolution predictions. Examine the use of CNN-based β-variational autoencoders for introducing stochasticity and improving interpretability in turbulent flow modeling. Investigate deep reinforcement learning approaches for flow control, including the control of 2D separation bubbles and opposition control in turbulent channel flow. This 50-minute presentation, hosted by Cambridge University Press, offers valuable insights into cutting-edge data-centric engineering techniques for researchers, engineers, and professionals interested in the intersection of data science and fluid dynamics.
Modelling and Controlling Turbulent Flows through Deep Learning
Cambridge University Press via YouTube
Overview
Syllabus
Intro
Modeling and controlling turbulent flow through deep learning
Motivation
Effect of Reynolds number for a given pressure gradient history: well-resolved LES
Adaptive simulations of NACA0012 profile with rounded wing tip
High-fidelity simulation of wing-tip vortex at Rec-200,000 and 5 degree angle of attack
Applications of machine learning to fluid mechanics
Outline of machine-learning applications to fluid mechanics
Flow reconstruction with a convolutional neu network (CNN)
CNN architecture
Turbulence statistics at Re,=550
Improving training performance: Transfer learning at Re,=180
Transfer learning from Re,=180 to 550
From sparse measurements to high-resolution predictions using GANS
FCN model for predictions closer to the wall
Self similarity in the overlap region: Off-wall boundary conditions
Turbulent flow in a simplified urban environment
CNN-based B-variational autoencoders CNN- Introducing stochasticity
Orthogonality: determinant of the cross-corre matrix
Effect of the penalization factor B
Optimality: ranking CNN-BVAE modes and interpretability
Deep reinforcement learning for flow control Introduction
Control of a 2D separation bubble
DRL and opposition control in turbulent channel flow: blowing and suction
Summary and Conclusions
Taught by
Cambridge University Press
