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 L17: Focus Session: Recent Advances in Data-driven and Machine Learning Methods for Turbulent Flows III |
Hide Abstracts |
Chair: Michael Brenner, Harvard University Room: 4c4 |
Monday, November 25, 2019 1:45PM - 1:58PM |
L17.00001: Machine learning meets mechanism: Mechanism of roll reversal in Rayleigh-B\'enard Convection Xi Chen, Xiaojue Zhu, Michael Brenner Understanding the precise physical events that underlie the reversal of circulation direction in Rayleigh-B\'enard (RB) convection has long been a mystery. We aim to solve this problem by using a machine learning model to classify the events related to a reversal in convection. We found a unique time scale indicating the transition between the two circulation directions of reversal, which is much shorter than the periodicity of reversal. We then try to “invert the neural network”, to discover the precise flow events that cause the physics on this timescale. This allows us to identify local patterns of the flow field that are critical to the reversal phenomena. We use these as the basis for building a phenomenological theory of reversal. We believe using interpretable machine learning in this way can be applied for the study of other fluid dynamics problems and even a lot of science problems. [Preview Abstract] |
Monday, November 25, 2019 1:58PM - 2:11PM |
L17.00002: Physics-Informed Echo State Networks for the Prediction of Extreme Events in Turbulent Shear Flows Nguyen Anh Khoa Doan, Wolfgang Polifke, Luca Magri A large number of turbulent flows exhibit extreme events. Extreme events are here defined as large-amplitude deterministic events, which suddenly occur aperiodically in the chaotic attractor. The time-accurate prediction of extreme events is challenging because of (i) the “butterfly effect”, which is the main property of the chaotic dynamics of turbulent flows, and (ii) the unpredictable nature of extreme events. We develop a physics-informed data-driven framework, the Physics-Informed Echo State Network (PI-ESN), to predict extreme events in turbulent flows. The PI-ESN consists of a reservoir of dynamical neurons, which learn the system’s dynamics from time-series of its evolution. We apply this method to a turbulent shear flow between two free-slip walls subject to a sinusoidal force. This flow displays abrupt transition between quasi-laminar and fully-turbulent states. We are able to time-accurately predict the flow evolution during these extreme events by imposing the physical principles as constraints in the learning algorithm. This physics-informed data-driven approach outperforms purely data-driven approaches, which opens up new horizons for the time-accurate prediction of turbulent flows by leveraging on data and physical principles. [Preview Abstract] |
Monday, November 25, 2019 2:11PM - 2:24PM |
L17.00003: Data-driven prediction of vortical structures in turbulent flows employing deep learning techniques Babak Kashir, Marco Ragone, Vitaliy Yurkiv, Farzad Mashayek The vortical structures are inherent characteristics of turbulent flows. Significant research has been conducted to understand, characterize and locate the vortical structures in turbulent flows in complex configurations. However, the identification of the size and location of the vortices in practical flows is often challenging due to the interplay of various parameters. To overcome these challenges, we have developed a deep learning model to identify and locate the vortical structures. The deep learning model is trained and tested in lid-driven cavity flow fields to predict the vortical structures in different conditions. The architecture of the model is characterized by multiple layers with random dropout and linear regularization, whereas the final prediction is performed through a binary classification. The second invariant of the velocity gradient tensor (known as the Q-criterion) is used to locate the vortical structures in fluid dynamics. This criterion describes a vortex as a continuous fluid region with the positive second invariant. The neural-network predictions are compared with the results from the previously validated numerical simulations. The present study allows for advance accelerated analysis of complex turbulent flows. [Preview Abstract] |
Monday, November 25, 2019 2:24PM - 2:37PM |
L17.00004: Flow Characteristics and Noise Performance on Side Mirror Models by 4D PTV and AI-Based Data Assimilation Kyung Chun Kim, Dong Kim, Mirae Kim, Edoardo Saredi, Fulvio Scarano A time-resolved three-dimensional Lagrangian Particle Tracking Velocimetry (4D PTV) has used to measure flow characteristics of three side mirror models adopting the Shake-the-box algorithm with four high-speed cameras on a robotic arm. Helium filled soap bubbles are used as tracers in the wind tunnel experiment to characterize flow structures around automobile side mirror models. Full volumetric velocity fields and evolution of vortex structures are obtained and analyzed. Instantaneous pressure fields are deduced by solving a Poisson equation based on the 4D PTV data. To increase spatial and temporal resolutions of velocity field, artificial intelligence (AI)-based data assimilation method has applied. ANFIS (Adaptive Neural Fuzzy Inference System) based machine learning algorithm works well to find hidden 3D vortical structures behind the automobile side mirror model. Using the high resolution ANFIS model, power spectrum of velocity fluctuations and sound level spectrum of pressure fluctuations are successfully obtained to assess flow and noise characteristics of side mirror models. [Preview Abstract] |
Monday, November 25, 2019 2:37PM - 2:50PM |
L17.00005: Deep learning the spanwise-averaged turbulent wake of a circular cylinder Bernat Font Garcia, Gabriel Weymouth, Vinh-Tan Nguyen, Owen Tutty Numerical simulations of long and flexible cylindrical structures become prohibitive at high Reynolds regimes because of the wide range of spatial and temporal scales that need to be resolved. We propose a new flow decomposition based on the spanwise average of the local three-dimensional (3D) strip which provides a two-dimensional formulation with additional statistical terms accounting for the 3D fluctuations. The latter unclosed terms are modelled through a convolutional neural network (CNN) trained on a high-fidelity dataset. The CNN is designed as a multiple-input multiple-output autoencoder inspired on image recognition architectures. The convolution operation ensures translational invariance and different inputs are tested aiming to provide a Galilean invariant model. \textit{A priori} results display 90\% correlation of the predicted turbulent fields and current work involves the \textit{a posteriori} analysis of the model plus the investigation of the model generalisation for different geometries and flow regimes. [Preview Abstract] |
Monday, November 25, 2019 2:50PM - 3:03PM |
L17.00006: Mechanisms of convolutional neural networks for learning three-dimensional unsteady wake flow Sangseung Lee, Donghyun You Recently, convolutional neural networks (CNNs) have been applied to predict or model flow dynamics. However, mechanisms of CNNs for learning flow dynamics are still not well understood, while such understanding is highly necessary to reduce trial-and-errors in designing networks. In the present study, we investigate the mechanisms of a CNN for prediction of three-dimensional unsteady wake flow behind a circular cylinder. Feature maps in the CNN are visualized to compare flow structures that the CNN extracts from flow at different flow regimes. A Fourier analysis is conducted to reveal the mechanisms, which enable the CNN to predict flow dynamics at different flow regimes, of a convolution layer to integrate and transport wave number information from flow. The integration and transportation characteristics of information of flow variables and histories in the CNN are discussed. [Preview Abstract] |
Monday, November 25, 2019 3:03PM - 3:16PM |
L17.00007: Learning effective viscosity for moderate Reynolds number Navier-Stokes equations Xiaojue Zhu, Michael Brenner We propose that with an appropriately chosen effective viscosity $\nu(Re)$, a linearization of the Navier-Stokes equations perfectly captures the drag-determining features of flows around unsteady translating bodies, in the Reynolds number of order hundreds. Two ways are implemented to find the effective viscosity. First, $\nu(Re)$ is determined so that the time-averaged shapes of separatrix, a fluid surface that delimits the compact region of fluid that is entrained by the moving point force, match as closely as possible between the linearization equation and the Navier-Stokes equation. Second, $\nu(Re)$ is learned from a data-driven method, i.e. a deep neural network, by minimizing the mean squared error loss. We compare the results for the two methods, and we find that the data-driven method dramatically outperforms the former traditional method. We apply the linearization to the classical problem of predicting vortex shedding around a cylinder. [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