Collaborative Research: Learning to estimate and control gust-induced aerodynamics

对于许多涉及升力表面的能源和推进系统，如风力涡轮机、固定翼和旋翼飞机以及涡轮机械，大振幅流体动力扰动或“阵风”是一个普遍存在的挑战。气流扰动通常是由大气引起的，由地形或天气引起，或由其他系统的空气动力学引入，如在风力发电场或一群飞行器中。当系统的重量和大小减小或天气事件变得更加极端时，它们就会变得相对更强。阵风会严重破坏系统的预期性能，或者在最坏的情况下，导致灾难性的故障。由于系统对阵风和驱动的空气动力学响应高度依赖于彼此，因此设计一种用于大振幅阵风缓解的自动化策略非常具有挑战性。强化学习（RL）是一种很有前途的方法来控制这种复杂的流体流动，它绕过了以前方法的许多障碍，但它受到训练负担的挑战：在naïve RL应用中，算法必须在训练期间看到适当范围的阵风条件和驱动响应，以确定每次遇到的最佳响应。如果算法包含流状态信息和流物理预测，则很可能加速强化学习的训练。利用流量状态信息增强RL的方法在很大程度上仍未被探索，主要是因为实际使用少量车载传感器实时推断这些信息的挑战。传感器只能提供有限的流量足迹，但这一足迹可以揭示大多数重要的流量信息。该项目将利用之前在非定常空气动力学计算和实验研究方面的工作，推进基于有限传感器的流动状态估计的最新技术，并缩小RL在流体动力学中实际应用的差距。该项目将通过实验和计算来估计固定翼或旋转叶片遭遇大振幅扰动时流动中的相干涡结构。通过使用详细的流量测量计算和实验，该计划将探索阵风遇到的各种关键流动物理，包括在广泛的雷诺数范围内的缩放效应，并研究在RL训练期间翼/叶片俯仰对这些遇到的影响。该程序将首次在实验室环境中演示阵风相互作用的强化学习控制。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Large-amplitude fluid dynamic disturbances, or “gusts”, are a pervasive challenge for many energy and propulsion systems involving lifting surfaces, such as wind turbines, fixed and rotary-wing aircraft, and turbomachinery. Flow disturbances are often atmospheric, caused by terrain or weather, or introduced by the aerodynamics of other systems, as in a wind farm or a swarm of air vehicles. They become relatively stronger as the system's weight and size decrease or as weather events become more extreme. Gust encounters can significantly undermine the desired performance of the system, or at worst, cause catastrophic failure. Devising an automated strategy for large-amplitude gust mitigation is exceptionally challenging because the aerodynamic responses of the system to the gust and to actuation are highly dependent upon each other. Reinforcement learning (RL) is a promising approach for control of such complex fluid flows that circumvents many of the obstacles to previous approaches, but it is challenged by the burden of training: in a naïve application of RL, the algorithm must see a suitably large range of gust conditions and actuation responses during training to determine the best response for each encounter.It is very likely that RL training can be accelerated if the algorithm incorporates flow state information and a prediction of flow physics. The augmentation of RL with flow state information remains largely unexplored, primarily because of the challenges of practically inferring this information in real time with a small number of on-board sensors. Sensors provide a limited footprint of the flow around them, but this footprint can reveal most of the essential flow information. This program will leverage prior work in computational and experimental investigations of unsteady aerodynamics to advance the state of the art of flow state estimation from limited sensors and to close the gap on practical use of RL in fluid dynamics. The program will deploy experiments and computations to estimate coherent vortex structures in a flow during encounters of a fixed wing or rotating blade with a large-amplitude disturbance. With use of both computations and experiments with detailed flow measurements, the program will explore a wide range of crucial flow physics in gust encounters, including scaling effects across a wide range of Reynolds numbers, and to study the influence of wing/blade pitching on these encounters during RL training. This program will demonstrate, for the first time, reinforcement learning control of gust interactions in a laboratory setting.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.