Shear-flow Prediction and Drag Estimation using AI-driven Methodology (SPADE-AIM)

• Applicant: Dr.-Ing. Gazi Hasanuzzaman
• Subject Area: Fluid Mechanics
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 569383977

Shear-driven turbulent flows dominate many indispensable but energy consuming engineering applications, where energy-saving flow control approaches can preserve environment by compensating fuel consumption. It is the friction parameters that plays a crucial role at least for incompressible turbulent boundary layers (TBL) that eventually contributes as half of the total energy consumption for aerodynamic applications such as subsonic passenger aircraft. Friction parameters are essentially the estimate of skin friction that is also used to express the turbulence statistics and scales. Measuring friction parameters accurately in TBL flows remains challenging using direct measurement approach, this becomes even more difficult at high Reynolds number. However, spatially and temporally under resolved data suffers the most in case of lack of data at the near-wall region. Particle Image Velocimetry (PIV) is known for its wide application in terms of simultaneously acquiring large volume of measurement data and eventually, dominates the turbulent research. However, common PIV applications suffers strongly within the viscous sub-layers (VSL) cannot resolve the friction parameters. Large fraction of the research community deploy other approaches in conjunction with PIV. Within this proposal, a novel predictive approach is outlined combining machine learning model and experimental data in order to predict the friction parameters. Physics-Informed-Neural-Networks (PINN) model will be incorporated with the PIV data for training in order to make out-of-the-field predictions of velocity data within a TBL configuration. This also includes different wall boundary conditions with wall-normal blowing at different rates. As the wall-normal blowing is well known drag control technique that can have long downstream persistent effects, recent evidence shows that they can also modify the large scale structures in terms of their energy. Deep neural network based PINNs have been proven effective predicting the mean flow fields in terms of velocity vectors as well as the Reynolds stresses. The model architecture uses two sets of points for supervised and unsupervised learning to train the network. Here, the governing laws such as Navier-Stokes equations are solved for unsupervised learning and their residuals are used as the unsupervised loss. On the other hand, supervised learning refers to the training process by computing a supervised loss for the data for which the targets/reference PIV data is available. Finally, the total loss is estimated by the sum of both these losses. Additionally, the ML model will be made robust with the validation of direct measurements of friction parameters and with different wall boundary conditions. This scientific proposal will help to achieve a reliable prediction method for friction parameters within TBL flows with the help of coarse measurement data.

• DFG Programme: Research Grants
• International Connection: Sweden
• Cooperation Partner: Professor Ricardo Vinuesa, Ph.D.