LES simulation of an airfoil for Machine Learning techniques
||LES simulation of an airfoil for Machine Learning techniques|
||SNIC Small Compute|
||Jerol Soibam <email@example.com>|
||2019-06-04 – 2020-07-01|
This project is a part of Future Energy Center (FEC) at Mälardalen University, Västerås. The project aims to implement machine learning (ML) techniques to solve fluid dynamics problems to reduce computation time. In order to train the machine learning algorithm, it requires sets of training data. Therefore, it is essential to create a database of the computational fluid dynamics (CFD) simulation to train the ML algorithms (Convolutional Neural Network). To set up the database, Large eddy simulation will be carried out of an airfoil for a different angle of attack and different input velocity using OpenFOAM CFD software. The result obtained from the CFD simulation will be used as the input for CNN to train the neural network. The goal here is to predict the lift coefficient, velocity, and pressure profile of the airfoil for a new angle of attack and input velocity which is not calculated using CFD solver. This project aims to understand how machine learning can be used to provide a generalized solution for fluid dynamics problems.
CNN has been widely used in the field of image recognition, as it can naturally take into account the spatial information of the data. This property of CNN is a major advantage in fluid mechanics application since CNN has the possibility to capture the flow behaviour in space.
The advantages of using Machine learning / artificial neural networks (ANN) is that they provide continuous predictions (i.e. for a range of an operating parameter) in one go, whereas the numerical and experimental approaches provide discrete results (for a specific value of the parameter). Furthermore, the computing time for the ANN is significantly lower when compared to conventional numerical simulation.
In order to execute this project, numerous CFD simulation (55 cases) has to be first run, to create the database, and running a LES simulation requires huge computing power. Therefore, there is a need for supercomputing power to perform these LES simulations.