Bulletin of the American Physical Society
77th Annual Meeting of the Division of Fluid Dynamics
Sunday–Tuesday, November 24–26, 2024; Salt Lake City, Utah
Session T12: Low-Order Modeling and Machine Learning in Fluid Dynamics: Flow Control |
Hide Abstracts |
Chair: Luning Sun, Lawrence Livermore National Laboratory Room: 155 B |
Monday, November 25, 2024 4:45PM - 4:58PM |
T12.00001: Comparative Analysis of Manifold Learning Techniques for Controlled flows Alicia Rodríguez-Asensio, Stefano Discetti, Andrea Ianiro Turbulent flows, despite their high dimensionality, exhibit recurrent patterns known as coherent structures, suggesting the possibility of representing key dynamics on a low-dimensional manifold. Manifold learning aims to identify such low-dimensional surfaces. Farzamnik et al. (2023, J Fluid Mech, 955:A34) demonstrated the effectiveness of low-dimensional manifold learning in describing shear flows. However, control inputs alter flow dynamics, complicating manifold identification. This work compares the performance of key data-driven manifold learning techniques for controlled flows, including standard and kernel Principal Component Analysis (kPCA), Multidimensional Scaling (MDS), Isometric Feature Mapping (ISOMAP), and Locally Linear Embedding (LLE). The test case is the fluidic pinball, a configuration of three cylinders in a uniform flow with independently controlled rotation. Despite its relatively simple dynamics, it offers a wide range of control possibilities, allowing an extensive span of different flow configurations. The results reveal that nonlinear methods capture meaningful coordinates in controlled flows, which can be used for low-order modeling. |
Monday, November 25, 2024 4:58PM - 5:11PM |
T12.00002: Actuation manifold from snapshot data Luigi MARRA, Guy Y. Cornejo Maceda, Andrea Meilán-Vila, Vanesa Guerrero, Salma Rashwan, Bernd R. Noack, Stefano Discetti, Andrea Ianiro Data-driven manifold learning has emerged as a promising technique for extracting low-dimensional representations from complex high-dimensional data. In this study, we propose a data-driven methodology to learn a low-dimensional manifold for controlled flows, referred to as an actuation manifold. |
Monday, November 25, 2024 5:11PM - 5:24PM |
T12.00003: Uncertainty quantification in low-order machine learning models of unsteady aerodynamics Hanieh Mousavi, Jeff D Eldredge In aerodynamics, accurately estimating the flow around airfoils is pivotal for effective flow control and optimal decision-making. To achieve this, we seek to use sensors on the airfoil to capture the underlying flow state. This study presents a data-driven approach to reconstructing the vorticity field and lift from limited surface pressure measurements using a non-linear lift-augmented neural network. As in some recent studies, our approach leverages deep learning tools to reduce the dimensionality of the system to a few latent variables, revealing that the inference of flow and lift from sensor data can be captured via a low-dimensional space. However, typical deep learning models do not inherently account for uncertainties, which is crucial for reliable predictions. Accordingly, we quantify two types of uncertainties: aleatoric uncertainty, arising from noisy measurements, and epistemic uncertainty, stemming from limitations and lack of knowledge in the mapping from pressure to the flow field. The uncertainty quantification provided by our model can be used to highlight the sensitivity of predictions from sensor measurements, strategically place sensors, and significantly improve decision-making processes in aerodynamic applications. |
Monday, November 25, 2024 5:24PM - 5:37PM |
T12.00004: Multi-fidelity Reinforcement Learning Control for Complex Dynamical Systems Luning Sun, Xin-yang Liu, Jian-Xun Wang, Siyan Zhao, Aditya Grover, Jayaraman Thiagarajan Controlling instabilities in complex dynamical systems is challenging in scientific and engineering applications. Deep reinforcement learning (DRL) has seen promising results for applications in different scientific applications. The many-query nature of control tasks requires multiple interactions with real environments of the underlying physics. However, it is usually sparse to collect from the experiments or expensive to simulate for complex dynamics. Alternatively, controlling surrogate modeling could mitigate the computational cost issue. However, a fast and accurate learning-based model by offline training makes it very hard to get accurate pointwise dynamics when the dynamics are chaotic. To bridge this gap, the current work proposes a multi-fidelity reinforcement learning (MFRL) framework that leverages differentiable hybrid models for control tasks, where a physics-based hybrid model is corrected by limited high-fidelity data. We also proposed a spectrum-based reward function for RL learning. The effect of the proposed framework is demonstrated on two complex dynamics in physics. The statistics of the MFRL control result match that computed from many-query evaluations of the high-fidelity environments and outperform other SOTA baselines. |
Monday, November 25, 2024 5:37PM - 5:50PM |
T12.00005: Efficient control of chaotic turbulent flow with reinforcement learning Sajeda Mokbel, Christian Lagemann, Esther Lagemann, Steve Brunton Chaotic systems often exhibit extreme events in which quantities-of-interest significantly deviate from the mean value, arbitrarily, for finite periods of time. Common examples include oceanic rogue waves, shocks in power grids, earthquakes, and turbulence. The detrimental effects of these systems make their control of utmost importance, however, finding the cause of these events and mitigating them remains a challenge. This work focuses on controlling the behavior of sinusoidally-driven turbulent flow, which exhibits extreme energy dissipation events due to non-linear energy transfers at different scales. Specifically, a model-free, low-dimensional reinforcement learning agent acts on selected energy modes to manipulate the system to achieve a desired behavior. The control goal of this work is two-fold: first, we demonstrate that mixing can be enhanced at low Reynolds numbers, and secondly, we illustrate empirically that the extreme events can be stabilized ahead of time. |
Monday, November 25, 2024 5:50PM - 6:03PM |
T12.00006: Optimal Riblet Spacing for Turbulent Flows Determined by Rank-1 Structured Input/Output Analysis Emma K Dufresne, Diganta Bhattacharjee, Mitul Luhar, Maziar S Hemati Patterned riblet surfaces have demonstrated the potential to passively reduce turbulent drag. Turbulent flows over riblets can be modelled and observed through structured input/output (I/O) analysis. Our work exploits recent theoretical advances to impose more precise nonlinear structure I/O analyses within structured I/O analysis. Our implementation reduces computational complexity by leveraging a rank-1 approximation of the linear dynamics, which then acts in feedback with a frozen flow field that results in maximum energy amplification of flow perturbations. Initial results suggest that imposing the more precise structure results in identification of instability mechanisms that are consistent with observations in prior direct numerical simulations. The optimal riblet spacing for a rectangular riblet geometry is identified for an incompressible channel flow at friction Reynolds number 180. The results are compared against predictions made using resolvent analysis and direct numerical simulations. |
Monday, November 25, 2024 6:03PM - 6:16PM |
T12.00007: Autonomous control of droplet generator for single and double droplets using Bayesian optimization Seongsu Cho, Haengyeong Kim, Seonghun Shin, Minki Lee, Jinkee Lee Droplet microfluidics is widely used in diverse applications, including functional particle fabrication, and biological assay. To achieve the desired result in application of droplet microfluidics, it is necessary to set the optimal flow rate. However, the optimal flow rate depends on multiple variables such as viscosity, channel dimension. Identifying optimal flow rate considering above factors is time-consuming and labor-intensive process. Previous studies have employed scaling laws or machine learning in an attempt to identify an optimal flow rate. However, these methods require a lot of experiment results. To overcome these limitations, we developed an autonomous control system which can control droplet generators for single and double droplets using Bayesian optimization. This system does not require huge training dataset and is applicable to droplet generating with various channel geometries and working fluids. Furthermore, we confirmed that it is applicable to not only single droplet generating but also double droplet generating. We believe these results can enhance accessibility of droplet microfluidics. |
Monday, November 25, 2024 6:16PM - 6:29PM |
T12.00008: Dynamic control and optimisationof plug-flow performanceusing machine learning Mosayeb Shams, Fuyue Liang, Nausheen Basha, Antonio Del Rio Chanona, Omar K. Matar Dynamic optimisation of transient performance in fluid systems is critical for various applications, such as heat exchange, and microfluidic mixing. However, traditional optimisation methods often rely on evaluating the entire performance profile over time, which can be computationally expensive and impractical. To address these limitations, we propose a novel approach that combines Bayesian Optimisation (BO) and Reinforcement Learning (RL) to efficiently optimise the transient performance of fluid systems based on early flow characteristics. Our key idea is to leverage the valuable information provided by Computational Fluid Dynamics (CFD) models in the BO framework, not as black-box functions, but as sources of local flow characteristics, such as cross-sectional velocity or mixing characteristics. By identifying the optimal early flow characteristics that strongly correlate with the desired performance, we can guide the optimisation process more efficiently than waiting for the entire performance profile to develop. The proposed BO-RL framework fills a critical gap in the current landscape of dynamic optimisation techniques for fluid systems and beyond, offering a computationally efficient and accurate solution for optimising transient performance based on early system characteristics. Furthermore, our approach is applicable to other domains facing similar dynamic optimisation challenges, such as energy grid management, where early characteristics like renewable energy generation and load profiles can predict and optimise system stability and efficiency. |
Monday, November 25, 2024 6:29PM - 6:42PM |
T12.00009: Flow control with latent dynamics model-based reinforcement learning Zhecheng Liu, Jeff D Eldredge Given the challenge of using classical control strategies for flow control due to the strong non-linearity and high dimensionality of fluid dynamics, Deep Reinforcement Learning has recently generated interest. Most applications thus far have used Model-Free Reinforcement Learning (MFRL) to train policies directly from CFD data. However, the intensive computational demand of MFRL, due to the high dimensionality of CFD, poses significant limitations in complex flow environments. To address this limitation, we propose a Model-Based Reinforcement Learning (MBRL) strategy, wherein the reduced model is trained from CFD data via two key tools. First, a Physics-Augmented Autoencoder learns to compress flow field snapshots to a very low dimensional latent space. Subsequently, a Latent Dynamics Model (LDM) learns to predict the dynamics in this space, thereby enabling accurate time-series forecasting of flow variables. We demonstrate the LDM's robustness and generalizability through accurate predictions in two distinct scenarios: a pitching airfoil in a highly disturbed environment and a Vertical-Axis Wind Turbine in a disturbance-free environment. We integrate the LDM into a MBRL framework applied to the disturbed airfoil scenario with the objective of minimizing the lift variation about a prescribed reference lift via pitch control. We show that our approach facilitates efficient policy learning within the latent space, significantly reducing the computational demand compared to MFRL. |
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. |
© 2025 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