Deep learning for non-intrusive sensing in turbulence with passive scalars

Abstract: The near-wall modelling of turbulent flows has been an active field of research due to the computational cost associated with the direct numerical simulations of such flow, which are characterized by a wide range of length and time scales. With the recent advancements in technological capabilities, the availability of high-fidelity data has enabled the construction of data-driven approaches to model turbulence. In this thesis, deep-learning models are used to model the dynamically important near-wall region in a turbulent boundary layer. As a first step, a direct numerical simulation (DNS) of an incompressible zero-pressure-gradient (ZPG) turbulent boundary layer (TBL) over a flat plate is performed using a pseudo-spectral code, SIMSON (Chevalier et al., 2007). The Reynolds number based on free-stream velocity and inlet displacement thickness is 450 and the passive scalars are simulated at Prandtl numbers of 1, 2, 4 and 6. Turbulence statistics for the flow and thermal fields are computed and compared against the numerical simulations at a similar Reynolds number. To generate the training, validation and test datasets for the neural network, the turbulent velocity fluctuation fields are sampled at various wall-normal locations, y+ = 15, 30, 50, 100 at a constant sampling time of ∆t+ = 0.99, in addition to the streamwise and spanwise wall-shear-stress fields, pressure field and heat flux fields at the wall. A fully convolutional network (FCN) based model is proposed for the prediction of two-dimensional velocity-fluctuation fields farther from the wall using the sampled fields at the wall. The quality of predictions from the network is assessed based on (i) the mean-squared error (MSE) between the predictions and the DNS fields, (ii) the relative percentage error in prediction of root-mean-squared (RMS) of fluctuations or fluctuation intensity and (iii) the correlation coefficient between the predicted and the DNS fields. Different types of predictions are performed, where the three components of the velocity-fluctuation fields are predicted simultaneously by the FCN, and these predictions are classified based on the input fields to the FCN. Three different types of predictions are presented in this study, and an auxiliary-loss-function approach is also introduced to improve the performance of the FCN. The results from the proposed data-driven model for ZPG TBL shows a good capability in the prediction of both the instantaneous fluctuation fields and the turbulent statistics like fluctuation intensity. In particular, the prediction of velocity-fluctuation fields at y+ = 30 using only the heat-flux field at Pr = 6 exhibits less than 12% error in the prediction of streamwise fluctuation intensity. The results obtained in this study indicate the potential of FCN in serving as a computationally effective tool to predict turbulent-velocity-fluctuation fields close to the wall using the inputs from the wall and finds useful application in flow-control problems.