Bulletin of the American Physical Society
76th Annual Meeting of the Division of Fluid Dynamics
Sunday–Tuesday, November 19–21, 2023; Washington, DC
Session R28: Emerging Trends in Fluid Dynamics: Physics-Informed Machine Learning and Dynamic Modeling |
Hide Abstracts |
Chair: Wai Tong Chung, Stanford University Room: 152A |
Monday, November 20, 2023 1:50PM - 2:03PM |
R28.00001: A Physics-Informed Machine Learning Approach for Predicting Atomized Drop Distributions in Liquid Jet Simulations Chris J Cundy, Shahab Mirjalili, Charlelie Laurent, Stefano Ermon, Ali Mani A key goal for simulations of liquid jet atomization is the accurate prediction of the size distribution and number density of atomized drops. The multi-scale nature of these flows makes it nearly impossible to capture all scales within a single simulation. Specifically, the breakup processes producing the smallest drops through secondary atomization often necessitate resolutions far below the Kolmogorov scale. Existing physics-based and stochastic breakup models fail to account for the local and instantaneous flow field and drop geometry. We present a physics-informed machine learning model for predicting the distribution of daughter drops generated during the breakup of under-resolved drops. We showcase proof-of-concept results from simplified configurations of 3D Taylor-Green vortex flows and homogeneous isotropic turbulence. By training on high-resolution simulations, the model can predict the result of breakup from severely under-resolved input fields. Compared to low-resolution simulations or phenomenological methods, our approach achieves superior accuracy in predicting drop size distribution and quantities of interest including surface area distribution and breakup probability. |
Monday, November 20, 2023 2:03PM - 2:16PM |
R28.00002: Physics-Informed Neural Networks (PINN) for Enhanced Dynamic Modeling and Reverse Problem Solving in an Electro-Wetting Operated Microfluid Prism Chihoon Song, Duck Gyu Lee, Jeongsu LEE, Keunhwan Park Physics-Informed Neural Networks (PINNs) are transforming the field of machine learning, offering powerful solutions to complex problems in fluid dynamics. This research introduces the application of PINNs to the intricate challenges associated with Electrowetting On Dielectric (EWOD) operated microliquid prisms, a task that has traditionally posed significant hurdles for conventional numerical methods. |
Monday, November 20, 2023 2:16PM - 2:29PM |
R28.00003: Predicting the three-dimensional separating flow in a diffuser using physics-informed neural networks Ben Steinfurth, Julien Weiss We study the capabilities of physics-informed neural networks (PINNs) that are trained to capture the three-dimensional mean flow of a turbulent separation bubble that occurs inside a one-sided diffuser. The network output is constrained by the three-dimensional Reynolds-averaged Navier-Stokes equations on the one hand and an extensive experimental database on the other. The latter contains the mean wall pressure field on the diffuser surface as well as approximately 730,000 three-component velocity vectors spanning the entire flow domain. In addition, the mean two-component wall shear-stress field acquired on the diffuser surface is employed to assimilate the velocity gradient at the wall. It is shown that measurement artefacts pertaining to the velocity field data are rectified thanks to the physics-informed approach. Furthermore, the velocity can be predicted reliably in the near-wall region where no measurement data are available. Here, a PINN trained with wall shear-stress data showed a better performance than an alternative model where this data was not provided. Leveraging the training data at the boundaries of the flow domain, we also demonstrate that the complex three-dimensional velocity field can be reconstructed with reasonable accuracy requiring as few as 0.01 % of the velocity training dataset (i.e., only 100 vectors). This gives rise to less elaborate acquisitions of three-dimensional flow fields where extensive velocity field measurements can be substituted with a few single-point measurements. |
Monday, November 20, 2023 2:29PM - 2:42PM |
R28.00004: DNS and physics-informed surrogate models of surfactant-laden dispersed flows Juan Pablo Valdes, Fuyue Liang, Lyes Kahouadji, Sibo Cheng, Seungwon Shin, Jalel Chergui, Damir Juric, Omar K Matar This study seeks to elucidate the fundamental physics governing surfactant-laden liquid-liquid dispersion processes under industrially relevant scenarios (i.e., static mixing). Using a DNS approach, we explore different surfactant physicochemical parameters (i.e., elasticity, desorption, and adsorption kinetics), where we compare relevant metrics (i.e., droplet count, size distribution) and interrelate them with the underlying physics captured in each case. We explicitly account for the role of Marangoni stresses during deformation and breakage. The rich data fields extracted from DNS are used to train surrogate models that can provide inexpensive, yet accurate, physics-informed predictions of key dispersion performance metrics calculated through DNS. We explore the application of deep convolutional recurrent autoencoders (CAE) to construct a low-dimensional representation of the dynamics obtained through DNS, and subsequently train neural networks with Long Short-Term Memory (LSTM) units to reconstruct the full physics and predict the dynamical evolution of metrics such as droplet count, size and local interfacial tension. |
Monday, November 20, 2023 2:42PM - 2:55PM |
R28.00005: Discovering self-similar blow-up solutions using physics-informed neural networks Yongji Wang, Ching-Yao Lai, Tristan Buckmaster, Javier Gomez Serrano One of the most challenging open questions in mathematical fluid dynamics is whether an inviscid incompressible fluid, described by the 3-dimensional Euler equations, with initially smooth velocity and finite energy can develop singularities (blow-ups) in finite time. This long-standing open problem is closely related to one of the seven Millennium Prize Problems which considers the Navier-Stokes equations, the viscous analogue to the Euler equations. In this talk, we present a novel numerical approach utilizing physics-informed neural networks (PINNs), that enables the discovery of self-similar blow-up solutions to various fluid equations, ranging from 1-D Burgers' equation to the 3-D Euler equations with a cylindrical boundary. Moreover, we introduce multi-stage neural networks that achieve machine precision accuracy for predicting blow-up solutions, forming the basis for rigorous computer-assisted proofs of them. This breakthrough sheds new light to the century-old mystery of capital importance in the field of mathematical fluid dynamics. |
Monday, November 20, 2023 2:55PM - 3:08PM |
R28.00006: Semi-supervised machine learning model for Lagrangian state estimation Reno Miura, Koji Fukagata In recent years, many studies have demonstrated the strength of supervised machine learning models for fluid state estimation. However, most of the studies assume that the sensors are fixed and that the high-resolution ground truth can be prepared. In practical situations, however, the sensors are not always fixed and may be floating. For example, in oceanography and river hydraulics, sensors are generally floating. Additionally, floating sensors make it more difficult to collect the high-resolution ground truth. We here propose a machine learning model for state estimation from such floating sensors without requiring high-resolution data for training. This model estimates velocity fields only from floating sensor measurements and is trained with a loss function using only sensor measurements and locations. We call this loss function as a "semi-supervised" loss function, since sensor measurements are used as the ground truth but high-resolution data of the entire velocity fields are not required. To demonstrate the model without high-resolution data for the training process, we consider two-dimensional decaying isotropic turbulence. Our results reveal that this model can estimate velocity fields with reasonable accuracy when the sensors are spatially dispersed to some extent in the domain. We also discuss the estimation accuracy dependence on machine learning methods, the number of sensors. |
Monday, November 20, 2023 3:08PM - 3:21PM |
R28.00007: Cooperative swimming at low Reynolds numbers using deep reinforcement learning Yangzhe Liu, Zonghao Zou, On Shun Pak, Alan C. H. Tsang Biological microswimmers perform cooperative behaviors to exploit their fluid environments, leading to enhanced swimming performance and/or energy efficiency. In this study, we employed a deep reinforcement learning approach to investigate the effective strategy for cooperative locomotion of reconfigurable swimmers. We consider a pair of 3-sphere swimmers arranged in a collinear configuration in a low Re fluid. The strategy obtained by the reinforcement learning approach consists of an approach stage and a synchronization stage. During the approach stage, the swimmers approach each other to reduce their relative distance and increase their hydrodynamic interactions. Subsequently, in the synchronization stage, the swimmers synchronize their locomotory gaits with a particular phase mismatch, resulting in locomotion performance that significantly surpasses an individual swimmer. This research highlights the potential of reinforcement learning in the control of cooperative behaviors of multiple swimmers. |
Monday, November 20, 2023 3:21PM - 3:34PM |
R28.00008: Reinforcement learning of reconfigurable microswimmers Alan C. H. Tsang, On Shun Pak The recent surge in applying artificial intelligence (AI) to fluid mechanics and propulsion problem has demonstrated the tremendous potential of AI in advancing solutions that are usually difficult to achieve with the traditional frameworks. In this talk, we will summarize our recent research in the reinforcement learning (RL) of reconfigurable microswimmers in a low Reynolds number fluid. We will discuss a set of simple models of reconfigurable microswimmers consisting of spheres with movable arms. We will introduce two RL approaches for a fully discrete reconfiguration system and a system with continuous variables, respectively. We will showcase how RL enables these swimmers to self-learn how to swim, rotate, and navigate towards specific targets. These findings present the initial step towards intelligent autonomous manipulation of artificial microswimmers. |
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