PECASE: Uncovering Nonlinear Flow Physics with Machine Learning Control and Sparse Modeling
Abstract
The modeling and control of fluid flows remains a grand challenge problem of the modern era, with potentially transformative scientific, technological, and industrial impact. Indeed, better understanding of complex flow physics may enable drag reduction, lift increase, mixing enhancement, and noise reduction in domains as diverse as transportation, energy, security and medicine. Fluid dynamics is a canonically difficult problem because of strong nonlinearity, high-dimensionality, and multi-scale physics; both modeling and control may be thought of as extremely challenging optimization problems. Recent advances in machine learning and sparse optimization are revolutionizing how we approach these traditionally intractable problems. The objective of the proposed work is to 1) use machine learning control (MLC) to push flows to new regimes and behaviors, and 2) use the recent sparse identification of nonlinear dynamics (SINDy) architecture to characterize the underlying flow physics, resulting in interpretable models. We envision that these methods will enable the discovery of novel flow physics as well as practical new control strategies to achieve improved performance in engineering flows. This work will combine MLC and SINDy for use in fluid dynamics to simultaneously explore and characterize new flow regimes. MLC constitutes a rapidly developing set of optimization techniques that may be applied to modify the behavior of strongly nonlinear flows; however, the resulting control laws are uninterpretable black-box expressions that yield limited insight into underlying physical mechanisms. The recent SINDy algorithm has been highly effective in producing efficient, interpretable, nonlinear reduced-order models of fluid flows; however, SINDy is only as good as the training data, and may yield an incomplete picture when data is sampled from an attractor without transients. Moreover, it may be a waste of resources to painstakingly characterize a model of a fluid attractor, just to apply control and have the dynamics change. Combining these methods will result in rich data to build models of the underlying flow physics, which will in turn improve the interpretability and performance of the controllers. In this work, fluid flows will be investigated with machine learning control and sparse model identification. These flows are chosen to exhibit interesting phenomena, including vortex shedding, dynamic stall, and broadband frequency cross-talk. In particular, this study will investigate 1) flow past a cylinder (laminar, 2D, simulated), 2) a new flow configuration consisting of three independently rotating cylinders (broadband, 2D, simulated), and 3) flow past a cross-flow turbine in a water channel (broadband, 3D, experimental). Applying machine learning control to these systems will explore new flow regimes, not only uncovering effective control strategies to manipulate the flows, but revealing new forced attractors, transients, and intermittent events. Further, sparse model identification will provide physically intuitive models, discovering essential nonlinear interaction terms in the dynamics. This data-driven framework has been shown capable of incorporating known physics and uncovering novel mechanisms, so that it will complement traditional modeling and control efforts. The additional time and resources of a PECASE award will facilitate a considerably more involved and ambitious effort, with additional work to 1) discover new generalized functions, or motifs, for the solutions of nonlinear fluid dynamic systems, 2) accelerate MLC and SINDy for real-time implementation, and 3) investigate the extreme flow control performance exhibited by insect flight in the new state-of-the-art UW hyper-sensed wind tunnel. At the end of the proposed work, we will have developed a framework to control and characterize fluids that improves with increasing data, positioning it to capitalize on the big data revolution.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 14, 2019
- Source ID
- W911NF1910045
Entities
People
- Steven Brunton
Organizations
- Army Contracting Command
- United States Army
- University of Washington