Bulletin of the American Physical Society
72nd Annual Meeting of the APS Division of Fluid Dynamics
Volume 64, Number 13
Saturday–Tuesday, November 23–26, 2019; Seattle, Washington
Session P17: Focus Session: Recent Advances in Data-Driven and Machine Learning Methods for Turbulent Flows IV |
Hide Abstracts |
Chair: Kevin Carlberg, Sandia National Laboratories Room: 4c4 |
Monday, November 25, 2019 5:16PM - 5:29PM |
P17.00001: Overview on sparsity in fluids Zhe Bai, Steven Brunton Fluid flows are typically represented in high-dimension, although they often exhibit low-dimensional patterns. Understanding these patterns and their evolving dynamics is crucial for control. Thus, discovering these patterns from experimental and numerical data is a central challenge in fluid dynamics. The existence of these low-dimensional flow patterns also enable efficient sensing strategies, sparsity-promoting optimization, and randomized methods in fluids. In this talk, we will discuss integrated sparse sensing and modal decomposition for fluids, which involve compressed sensing, proper orthogonal decomposition, and dynamic mode decomposition, providing a foundation for pattern recognition and low-rank structure discovery of high-dimensional systems. These data-driven models save tremendous online experimental and computational resources by leveraging the existence of patterns. We will illustrate these ideas on a variety of engineering applications. [Preview Abstract] |
Monday, November 25, 2019 5:29PM - 5:42PM |
P17.00002: Deep Neural Networks for Reduced Order Models for Fluid Flows William Wolf, Hugo Lui We present two numerical methodologies for construction of reduced order models, ROMs, of fluid flows through deep neural networks, DNNs. Here, the neural networks are used for regression and the frameworks are implemented in two contexts: one employs deep feedforward neural networks using a procedure similar to the sparse identification of non-linear dynamics algorithm, SINDy, and another is implemented using convolutional neural networks directly to the flow snapshots. The methods are tested on the reconstruction of a turbulent flow computed by a large eddy simulation of a plunging airfoil under dynamic stall. The reduced order models are able to capture the most energetic dynamics of dynamic stall including the leading edge stall vortex and the subsequent trailing edge vortex. The numerical framework allows the prediction of the flowfield beyond the training window and we demonstrate the robustness of the current ROMs constructed via deep neural networks through a comparison with sparse regression. The DNN approaches are able to learn transient features of the flow and present more accurate and stable long-term predictions compared to sparse regression. [Preview Abstract] |
Monday, November 25, 2019 5:42PM - 5:55PM |
P17.00003: Modeling particle-induced turbulence using sparse regression with embedded invariance Sarah Beetham, Jesse Capecelatro Turbulence is ubiquitous in science and industry and is nearly always multiphase. Given current computational capabilities and the wide range of time- and length-scales of industrial systems, direct numerical simulation (DNS) is prohibitively costly. Thus, some degree of modeling must be employed. Current state-of-the-art modeling for turbulent multiphase flows is predominantly based on extensions to single-phase models, making it largely unsuccessful beyond the dilute limit. This eliminates the augmentation of existing models as an option for solving the multiphase closure problem. Our goal is to propose compact, tractable multiphase turbulence closures. Thus, we first derive the exact Reynolds Stress equations for multiphase flows, which identifies the specific terms requiring modeling. To arrive at these closures, we use sparse regression with embedded invariance. In this talk, we demonstrate the promise of this technique for three classes of flow with increasing difficulty: (1) single-phase free shear turbulence, (2) turbulent flow over periodic hills and (3) gas-solid sedimenting flow. [Preview Abstract] |
Monday, November 25, 2019 5:55PM - 6:08PM |
P17.00004: Leveraging Dynamics for Near-Optimal, Ultra-Sparse Sensor Placement Samuel Otto, Clarence Rowley Optimal sensor placement in high-dimensional nonlinear dynamical systems like fluid flows remains a challenging problem. Most current methods identify an overly large number of sensors because they do not make use of the time histories at each sensor location. Our work begins by constructing a POD subspace capturing the finite-time state trajectories of interest. The sensors must be able to robustly reconstruct trajectories in this subspace, leading to an objective function that has ``nice'' mathematical properties (namely, it is normalized, monotone, and submodular). These properties guarantee that an accelerated greedy algorithm for sensor placement has performance within a constant factor of the optimal performance. In addition to reconstructing trajectories in POD subspaces, our method can be extended to identify even fewer sensors that enable nonlinear reconstruction of trajectories on curved manifolds (which we call ``ultra-sparse'' sensor placement). We illustrate these methods with examples including a cylinder wake flow and Burgers equation. [Preview Abstract] |
Monday, November 25, 2019 6:08PM - 6:21PM |
P17.00005: Space-time recovery of high-resolution turbulent flow fields with machine learning based super resolution Kai Fukami, Koji Fukagata, Kunihiko Taira In recent years, the use of machine learning based super-resolution analysis has enabled accurate reconstruction of high-resolution image from its low-resolution counterpart. Moreover, machine learning techniques referred to as inbetweening have also been developed to estimate data in between temporal snapshots. Here, we combine two of these approaches to reconstruct complex multi-scale turbulent flows both in space and time. A convolutional neural network-based architecture called hybrid Downsampled Skip-Connection and Multi-Scale (DSC/MS) model is developed for the recovery of complex flow fields. The proposed model is applied to two-dimensional isotropic turbulence and three-dimensional turbulent channel flow at $Re_{\tau}$ = 180 so as to demonstrate its capability in reconstructing spatio-temporal high-resolution turbulent flow fields from their coarse flow field data. We find that the present approach is able to accurately recover the high-resolution flow fields with only a modest amount of training data, despite the turbulent flow being complex and multi-scale in nature. The first two authors acknowledge the support by JSPS (18H03758). The last author thanks the support from ARO (W911NF-17-1-0118), and AFOSR (FA9550-16-1-0650). [Preview Abstract] |
Monday, November 25, 2019 6:21PM - 6:34PM |
P17.00006: Reinforcement learning versus linear control of Rayleigh-Bénard convection Alessandro Corbetta, Gerben Beintema, Luca Biferale, Pinaki Kumar, Federico Toschi Thermally driven turbulent flows are common in nature and in industrial applications. The presence of a (turbulent) flow can greatly enhance the heat transfer with respect to its conductive value. It is therefore extremely important -in fundamental and applied perspective- to understand if and how it is possible to control the heat transfer in thermally driven flows.\\ In this work, we aim at maintaining a Rayleigh–Bénard convection (RBC) cell in its conductive state beyond the critical Rayleigh number for the onset of convection. We specifically consider controls based on local modifications of the boundary temperature (fluctuations). We take advantage of recent developments in Artificial Intelligence and Reinforcement Learning (RL) to find -automatically- efficient non-linear control strategies. We train RL agents via parallel, GPU-based, 2D lattice Boltzmann simulations. Trained RL agents are capable of increasing the critical Rayleigh number of a factor 3 in comparison with state-of-the-art linear control approaches. Moreover, we observe that control agents are able to significantly reduce the convective flow also when the conductive state is unobtainable. This is achieved by finding and inducing complex flow fields. [Preview Abstract] |
Monday, November 25, 2019 6:34PM - 6:47PM |
P17.00007: Active feedback control of flow over a circular cylinder with wall pressure sensor using machine learning Jinhyeok Yun, Jungil Lee In the present study, we conduct active feedback control of laminar and turbulent flows over a circular cylinder with wall pressure sensor for suppression of vortex shedding in the wake. The blowing and suction actuations are imposed at the wall before the flow separation, and their magnitudes are proportional to the transverse velocity in the the wake. To avoid direct measurement of velocity in the wake, we build an artificial neural network (ANN) between the pressures on the cylinder surface and the transverse velocities. For the learning process to build ANN, instantaneous flow data sets are obtained from numerical simulations of flow over a cylinder at $Re =$ 60 and 3900. The performance of ANN is assessed with the locations of wall pressures, structures of neural network, and etc. It is found that the wall pressures on the cylinder surface can accurately predict velocities in the wake with the neural network built. Active feedback control combined with this neural network successfully suppresses the vortex shedding behind the cylinder, leading to reductions of the drag and lift fluctuations of cylinder. [Preview Abstract] |
Monday, November 25, 2019 6:47PM - 7:00PM |
P17.00008: Deep Learning for In-situ Compression of Large CFD Simulations Ryan King, Andrew Glaws, Michael Sprague The ExaWind project seeks to develop blade-resolved LES simulations of wind turbines for next-generation exascale computing architectures. Such simulations are expected to generate data with over a billion degrees of freedom and upwards of a million time steps, requiring significant computational resources to be dedicated to data storage, visualization, and analysis. In many cases, performing these tasks on the full dataset is intractable, prompting the need for in-situ data compression. The singular value decomposition (SVD) is the standard matrix compression approach; however, the linear nature of the low-rank approximation limits its ability to reconstruct highly nonlinear turbulent flow data. In this work, we explore deep learning methods for in-situ data compression, specifically a deep convolution autoencoder network that that maps 3D turbulent fields to a low-dimensional latent space. We compare the autoencoder against single-pass randomized SVD approaches in lossy restart studies where simulations are check-pointed and restarted. Our results show that an autoencoder trained on canonical turbulent flows can be applied to unseen configurations and is competitive with a single-pass SVD in terms of compression ratio, error, and computational cost. [Preview Abstract] |
Follow Us |
Engage
Become an APS Member |
My APS
Renew Membership |
Information for |
About APSThe American Physical Society (APS) is a non-profit membership organization working to advance the knowledge of physics. |
© 2024 American Physical Society
| All rights reserved | Terms of Use
| Contact Us
Headquarters
1 Physics Ellipse, College Park, MD 20740-3844
(301) 209-3200
Editorial Office
100 Motor Pkwy, Suite 110, Hauppauge, NY 11788
(631) 591-4000
Office of Public Affairs
529 14th St NW, Suite 1050, Washington, D.C. 20045-2001
(202) 662-8700