Evaluation of an in-house GPU based CFD solver

dc.contributor.authorJohansson, Christoffer
dc.contributor.departmentChalmers tekniska högskola / Institutionen för mekanik och maritima vetenskapersv
dc.contributor.departmentChalmers University of Technology / Department of Mechanics and Maritime Sciencesen
dc.description.abstractThis thesis describes the evaluation of an in-house CFD solver at GKN aerospace. The solver is called CUDA since it is GPU based, and CUDA cores are thereby used. Legacy turbine rear structures, designs from previous projects, are used when evaluating the solver. The first part of this thesis compares two different versions of CUDA. A correction in the definition of axial wall shear stress makes the new version more correct. By postprocessing flow parameters directly from the solution the new version is 25% faster than the previous version. The new version was officially released as a result of the validations performed in this thesis. The second part of this thesis evaluates the CUDA solver by comparing CFD results with the commercial CFD solver Fluent. CUDA shows similar trends as fully resolved k-ε realizable in Fluent. CUDA uses k-ε with realizability limiters as well, but for the wall treatments wall-functions are used. One evaluated flow parameter is total pressure loss. The pressure loss is presented in three different ways with respect to the inlet swirl angle, using a so called loss bucket and also two factors with normalized result. The two normalized factors are called off-design factor and loss difference. For all these three ways of presenting the pressure loss CUDA and Fluent predicts similar trends for all turbine rear structures. This correlation in pressure loss is predicted until separation occurs. Separation occurs later for CUDA than Fluent. For CUDA the point of separation is predicted to occur in average at 4 degrees more swirl than fully resolved k-ε realizable and 11.5 degrees later than k-ω SST for the turbine rear structures. The pressure loss is predicted to be lower in CUDA. CUDA predicts between 1% and 13% lower pressure loss than fully resolved k-ε realizable in Fluent, in average 6.75%.
dc.relation.ispartofseriesExamensarbete - Institutionen för mekanik och maritima vetenskaper : 2018:12
dc.subjectHållbar utveckling
dc.subjectStrömningsmekanik och akustik
dc.subjectSustainable Development
dc.subjectMechanical Engineering
dc.subjectFluid Mechanics and Acoustics
dc.titleEvaluation of an in-house GPU based CFD solver
dc.type.degreeExamensarbete för masterexamensv
dc.type.degreeMaster Thesisen
local.programmeApplied mechanics (MPAME), MSc
Ladda ner
Original bundle
Visar 1 - 1 av 1
Bild (thumbnail)
4.05 MB
Adobe Portable Document Format