The neural network fluid dynamicist: networks in feedback for flow control, sensor placement, and flow physics analysis

This talk will discuss a methodology for exploiting neural network architectures to perform a variety of common flow analysis and control tasks. Starting with a nonlinear fluid system equipped with some means of actuation, we first identify a neural network surrogate model for the actuated system. We then use this surrogate to train a second neural network model, designed to achieve a desired control objective. Through an iterative training process for both the model and controller neural networks, we obtain feedback control laws designed to drive unstable, nonlinear systems to their equilibria. Through the application of L1 regularization within the neural network loss function, optimal sensor locations can be identified from a larger set of candidates, allowing for stabilization using a reduced set of real-time measurements. This methodology is validated on several nonlinear systems, including a modified Kuramoto-Sivashinsky equation, spanwise-constant channel flow, and confined cylinder flow. We further show that identified neural-network models can be exploited to find not only unstable equilibria, but also leading linear stability eigenmodes. This demonstrates that rather than just being black-boxes, such neural network models can reveal pertinent flow physics.