Development of a high-order method for the prediction of low-pressure turbine losses using quasi-direct numerical simulations
- Autores: Marc Bolinches i Gisbert
- Directores de la Tesis: Roque Corral García (dir. tes.)
- Lectura: En la Universidad Politécnica de Madrid ( España ) en 2020
- Idioma: español
- Materias:
- Texto completo no disponible (Saber más ...)
- Resumen
In this thesis the implementation of a compact high-order Computational Fluid Dynamics (CFD) method and its validations is tackled. The chosen compact scheme is that of the Flux-Reconstruction (FR) method. This method is indeed a family of methods that can recover other schemes such as Discontinuous Galerkin method. These methods are particularly well suited for massively parallel hardware architectures since many operation per degree of freedom are performed which are independent of neighboring cells. The implementation of the CFD method has been made on utility Graphical Processor Units (GPU) in order for the software to take advantage of their parallel architecture.
The GPUs used in this work are single precision hardware. This imposed further action in order to keep the precision of the solution. Particularly, action was needed in the implementation of our time-marching algorithm. This was mainly due to the fact that very small time-steps were imposed by stability conditions. The Kahan summation algorithm was used in order to prevent single-precision errors from appearing as the solution marched in time.
The validation of the method was carried out against a very rich array of high-quality experimental data. This validation case consisted of a low velocity Low Pressure Turbine linear cascade designed by ITP Aero. The experiments were carried out at the Polytechnic University of Madrid. Pressure distribution were measured over the profile; velocity and root-mean-square (RMS) of velocity fluctuations were measured in the boundary layer; and total pressure, velocity and RMS of velocity and angle distributions were measured at the exit measurement station. All these data were compared against the numerical results with outstanding agreement for all the cases. Global quantities such as the cascade turning angle and cascade losses also agreed well with experimental data. More importantly, the Reynolds lapse rate, i.e., the variation of the cascade loss with Reynolds number, was captured by our simulations. This thorough comparison is expected to increase confidence in compact high-order schemes.
