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 L20: Low-Order Modeling and Machine Learning |
Hide Abstracts |
Chair: Rajeev Jaiman, University of British Columbia Room: 602 |
Monday, November 25, 2019 1:45PM - 1:58PM |
L20.00001: Using Deep Neural Networks for Data-Driven Prediction of Fluid Forces on Aerofoils Tharindu Miyanawala, Pasan Henadeera, Nalaka Samaraweera, Rajeev Jaiman We present an efficient deep learning technique for the prediction of fluid forces on aerofoils. The proposed technique relies on Convolutional Neural Network (CNN) technique. The aim is to predict the fluid forces for different aerofoil shapes at different angles of attack. The convolution with nonlinear rectification is employed to approximate the mapping between the aerofoil shape and the fluid forces. The deep neural network is fed by the Euclidean distance function and inviscid flow fields as the input and the target data generated by the XFOIL software and the full-order CFD computations for 525 aerofoil cases. The CNN is iteratively trained using the stochastic gradient descent method to predict the forces of different geometries and the results are compared with the full-order computations. A systematic convergence and sensitivity study is performed to identify the best dimensions of the deep-learned CNN. Within the error threshold, the prediction based on CNN got a speed-up an order of magnitude compared to the CFD results and consumes a fraction of computational resources. The proposed CNN-based approximation procedure will have a profound impact on the parametric design of aerofoils. [Preview Abstract] |
(Author Not Attending)
|
L20.00002: Airfoil Shape Optimization using Deep Q - Network Siddharth Rout, Prof. Chao-An Lin The feasibility of using reinforcement learning for airfoil shape optimization is explored. Deep Q-Network (DQN) is used over Markov's decision process to find the optimal shape by learning the best changes to the initial shape for achieving the required goal. The airfoil profile is generated using Bezier control points to reduce the number of control variables. The changes in position of control points is restricted to the direction normal to the chordline so as to reduce the complexity of optimization. The process is designed as a search for an episode of change done to each control point of a profile. The DQN essentially learns the episode of best changes by updating the temporal difference of Bellman Optimality Equation. The drag and lift coefficients are calculated from the distribution of pressure coefficient along the profile computed using XFoil potential flow solver. These coefficients are used to give reward to every change during learning process where the ultimate aim stands to maximize the cumulate reward of an episode. [Preview Abstract] |
Monday, November 25, 2019 2:11PM - 2:24PM |
L20.00003: Bringing Computational Fluid Dynamics at the heart of industrial processes: can Machine Learning help? Christos Varsakelis, Sandrine Dessoy Advances in both physical understanding and computational power have rendered several industrial problems amenable to a Computational Fluid Dynamics (CFD) analysis. However, even though in silico prototyping is gradually becoming the norm, utilizing the predictive power of CFD for real-time tasks, e.g. controlling a process, remains prohibitive. Recent literature has suggested that Machine Learning (ML) algorithms may be trained by CFD simulations and, subsequently, replace CFD codes due to their speed advantage. In this talk, we systematically evaluate this proposal through a series of industrial test cases of varying complexity. Both laminar and turbulent test cases are examined for various spatio-temporal scales. The talk concludes with the ranking of the performance of ML algorithms in terms of accuracy, size of training data required, speed of execution and computational cost. [Preview Abstract] |
Monday, November 25, 2019 2:24PM - 2:37PM |
L20.00004: Prediction of aerodynamic loads in turbulent flow conditions Andreas Natsis Birds and bats have a remarkable ability to isolate body motion in turbulent air. Hairs and feathers, spread out across their wings and body, sense the movement of the air (turbulence) before it has a significant effect on the motion of the animal. Analogously, distributed pressure sensors over a wing are used and the acquired data is analyzed with Neural Networks. An airfoil (NACA 0018, 60 cm span, 10 cm chord) with one degree of freedom (roll) was subjected to airflow with high intensity turbulence and an average speed of 10 m/s. The wing was tapped with multiple MEM pressure sensors with 1KHz sampling rate and its roll was recorded. Long short-term memory (LSTM) neural networks processed the information gathered by the pressure sensors and predicted roll by 20 ms. This is in stark contrast to most stability controllers utilized in flying vehicles to date. Current controllers make use of inertial measurement units (IMUs) located in the main body and require motion to occur before attempting to counter it. These results indicate that a bioinspired controller using pressure sensors is possible and may overcome inherent limitations of traditional IMU based controllers. [Preview Abstract] |
Monday, November 25, 2019 2:37PM - 2:50PM |
L20.00005: Airfoil control with Proximal Policy Optimization Denis Dumoulin, Philippe Chatelain Airfoil control generally relies on techniques based on a dynamical model of the actual system one wants to control. Such methods are in some cases limited by the expressiveness of the model and its linearization around equilibrium points. This is the case in aerodynamics where non-linearities (e.g., turbulent structures or dynamic stall) strongly affect the flight conditions and more specifically aerodynamic loads undergone by the flying body. To avoid such limitations, we propose the use of a model-free control method based on Reinforcement Learning (RL) algorithms. This study relies on a low-order inviscid planar solver which accounts for separation; this model controls its computational cost through lumping of the shed vortices in the wake. The complexity of the model is critical for RL methods which require large amounts of data. We then use Proximal Policy Optimization (PPO) to control the dynamics of a NACA0012 airfoil. This planar body goes through a random train of vortical dipoles while trying to keep its aerodynamic coefficients constant. Within a few thousands of episodes of training, the controller is able to stabilize the airfoil regardless of the incoming vortices using pressure measurements distributed on the body and its kinematics. [Preview Abstract] |
Monday, November 25, 2019 2:50PM - 3:03PM |
L20.00006: A reduced order model for store separation in high speed flow Nicholas Peters, John Ekaterinaris Store separation from aircraft and spacecraft has historically been a critical issue for the aerospace industry. Given the severity of the problem much effort has been spent on the topic. Yet, the process for identifying potential failures is a resource intensive and iterative process. A potential remedy for reducing iterations in this process is the implementation of an active controller for use during separation. The objective of this study is to design such a controller, using a reduced ordered model (ROM) for the flow around an external store undergoing separation. A combination of computational fluid dynamic (CFD) solvers, fun3D and Ansys Fluent, is employed to obtain flow fields around the vehicle and the store. Preliminary validation of the numerical results is initially carried out. Representative cases of store separation are obtained next. Proper Orthogonal Decomposition (POD) and Dynamic Mode Decomposition (DMD) are used to obtain leading modes that will be used to reconstruct a ROM of the flow field. These models will be compared with the objective of seeing which method better represents the numerical solution to the problem at different flow speeds with and without the presence of shock waves. The ROM will then be used to construct the controller. [Preview Abstract] |
Monday, November 25, 2019 3:03PM - 3:16PM |
L20.00007: Fast potential flow computations for low-order aerodynamic modelling Diederik Beckers, Jeff D. Eldredge Lightweight aircraft are vulnerable to flow separation induced by gusts. For purposes of regulating flight in the presence of such gusts, it is important to estimate the flow behavior and the instantaneous aerodynamic forces. In previous work, it was shown (D. Darakananda et al., Phys. Rev. Fluids 3, 124701, 2018) that low-order vortex models can be assimilated with sensor measurements to achieve this estimation. However, traditional vortex element models using Biot-Savart interactions can be computationally inefficient, particularly for 3D models. This work addresses grid-based computations for potential flows in 2D and 3D with the goal of implementing a low-order vortex model for fast modeling of separated aerodynamic flows and gust interactions. The immersed boundary projection method is used to solve for the vector potential field subject to the constraints introduced by the presence of a body. The equations are discretized on a staggered Cartesian grid and solved using the lattice Green's function. The accuracy of these computations is demonstrated for singular vortex elements in 2D and the extension to 3D flows will be discussed. [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