Adjoint-based machine learning for active flow control
Xuemin Liu, Jonathan F. MacArt
Abstract
We develop neural-network active flow controllers through a deep learning PDE augmentation method (DPM). In two-dimensional, incompressible, confined cylinder flow with Re = 100, we compare drag-reduction performance and optimization cost of adjoint-based controllers and deep reinforcement learning (DRL)-based controllers. The DRL-based controller demands 4,229 times the model complexity of the DPM-based one. The DPM-based controller is 4.85 times more effective and 63.2 times less computationally intensive to train than the DRL-based counterpart. In laminar compressible flows, successful extrapolation of the controller to out-of-sample flows demonstrates the robustness of the learning approach.
Topics & Concepts
Control theory (sociology)Computer scienceRobustness (evolution)ExtrapolationArtificial neural networkReinforcement learningDragLaminar flowCompressible flowControl engineeringCompressibilityArtificial intelligenceMathematicsEngineeringControl (management)Aerospace engineeringBiochemistryMathematical analysisChemistryGeneModel Reduction and Neural NetworksFluid Dynamics and Turbulent FlowsLattice Boltzmann Simulation Studies