The objective of this project is to develop numerical methods based on deep neural network (DNN) architecture and artificial intelligence. The DNN:s will be used to tackle three different problems,
1. To predict the wall-parallel velocity field in an open turbulent channel flow using the shear stress and pressure at the bottom of the channel as input data.
2. To predict the temporal evolution of the streamwise velocity component in a turbulent boundary layer using the wall pressure at several temporal instances.
3. To predict the wall heat flux in a Turbulent open-channel flow at different Prandtl numbers using velocity fields at different wall normal distances and temperature fields at the corresponding Prandtl number.
The PI and his team have investigated Task 1 and 2 using several different DNN architectures such as a fully convolutional network (FCN), a generative adversarial network (GAN), long short-term memory (LSTM). The architecture that we want to use in this project is the Transformer network. Just like recurrent neural networks (RNN:s) and LSTM the transformer was originally designed to process sequential input data. However the architecture can also be used in computer vision but in the form of the vision transformer (ViT). In this project the Transformer will be used to tackle problems 1-3. The performance of previously implemented architectures such as the FCN and GAN:s will be used as a baseline case where the expectation is that the Transformer will not only be able to achieve the same level of performance but also enhance the resolution of the predictions and decrease the error in the prediction.
The analysis will be performed based on four databases, three of which have been generated through direct numerical simulation (DNS) and one experimentally measured,
1. A turbulent open-channel flow at a specific friction Reynolds number.
2. An incompressible zero-pressure-gradient turbulent boundary layer over a flat plate.
3. A turbulent open-channel flow with specific friction Reynolds numbers at a range of various Prandtl numbers.
4. Experimental measurements of wall pressure and streamwise velocity from a wind tunnel.
Database 1 will be used for the analysis of problem 1, database 2 and 4 will be used for problem 2 and database 3 will be used for the analysis of problem 3.
The proposed method has excellent potential for closed-loop control applications relying on non-intrusive sensing and this is also the main motivation to carry out this project.