Abstract
In this paper we consider the problem of model-based learning of turbulentflows using mobile robots. The key idea is to use empirical data to improve onnumerical estimates of time-averaged flow properties that can be obtained usingReynolds-Averaged Navier Stokes (RANS) models. RANS models are computationallyefficient and provide global knowledge of the flow but they also rely onsimplifying assumptions and require experimental validation. In this paper, weinstead construct statistical models of the flow properties using GaussianProcesses (GPs) and rely on the numerical solutions obtained from RANS modelsto inform their mean. We then utilize Bayesian inference to incorporateempirical measurements of the flow into these GPs, specifically, measurementsof the time-averaged velocity and turbulent intensity fields. Our formulationaccounts for model ambiguity and parameter uncertainty via hierarchical modelselection. Moreover, it accounts for measurement noise by systematicallyincorporating it in the GP models. To obtain the velocity and turbulentintensity measurements, we design a cost-effective mobile robot sensor thatcollects and analyzes instantaneous velocity readings. We control this mobilerobot through a sequence of waypoints that maximize the information content ofthe corresponding measurements. The end result is a posterior distribution ofthe flow field that better approximates the real flow and also quantifies theuncertainty in the flow properties. We present experimental results thatdemonstrate considerable improvement in the prediction of the flow propertiescompared to pure numerical simulations.