Seeing the Wind: Leveraging flow-structure interactions for visual anemometry

点击翻译按钮获取中文摘要

Accurate measurements of wind speeds are crucial for many engineering applications, including wind energy, weather forecasting, drone navigation, and mitigation of dangers such as air pollution, wildfires, and airborne pathogens. Most wind-measuring devices used today are fixed in space and have high deployment costs. However, the built environment is already naturally instrumented with a variety of structures that move in response to the wind, from swaying trees to flapping flags. The principal aim of this project is to develop a physics-based understanding of those flow-structure interactions and to combine that knowledge with machine learning techniques to achieve unprecedented wind measurement capabilities. The project will incorporate laboratory measurements of vegetation exposed to controlled wind conditions in a large wind tunnel, and also field measurements in naturally occurring winds. K-12 engagement will be facilitated by emphasizing the timely applications of this research, especially firefighting and pollution monitoring, both of which are pressing challenges in southern California where the students live. An initial proof-of-concept has demonstrated the feasibility of using visual measurements of flow-structure interactions of vegetation to resolve wind speeds using a neural network. The goal of this project is to enable generalization beyond the initial training data set by incorporating a physics-based understanding of flow-structure interactions into the network architecture. The first objective is to determine which physical properties of the flow-structure interaction can be extracted by the model and are necessary for accurate wind speed predictions. Then, a broader training set and physics-based constraints will enable wind speed inference from classes of vegetation not used to train the model. A formal post hoc analysis of the neural network will further elucidate the salient flow physics. This project can improve the physical understanding of flow-structure interactions through development and subsequent analysis of the physics-based, machine learning approach to visual anemometry. The physical insights that will be gained from the data-driven approach will be considered along with a purely physics-based, first-principles approach to generalize the model and better understand its limitations. Moreover, because physics-informed machine learning is of broad interest in basic research, the approach to visual anemometry in this project can provide a template for similar efforts toward other research questions. A main outcome of this project, in addition to new knowledge regarding fluid dynamics and development of physics-informed machine learning, will be creation of a robust, quantitative visual anemometry technique.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.