CBL - Campus del Baix Llobregat

Projecte llegit

Títol: Desarrollo y validación de un solver CFD bidimensional acelerado por GPU y aplicación a la optimización evolutiva de perfiles alares

Estudiants que han llegit aquest projecte:

Director/a: MELLIBOVSKY ELSTEIN, FERNANDO PABLO

Departament: FIS

Títol: Desarrollo y validación de un solver CFD bidimensional acelerado por GPU y aplicación a la optimización evolutiva de perfiles alares

Data inici oferta: 01-02-2026     Data finalització oferta: 01-10-2026


Estudis d'assignació del projecte:
    GR ENG SIST AEROESP
Tipus: Individual
 
Lloc de realització: EETAC
 
Paraules clau:
CFD, Programación, Dinámica de Fluidos, Optimización, Algoritmo Genético
 
Descripció del contingut i pla d'activitats:
 
Overview (resum en anglès):
This work addresses airfoil optimisation as a search problem over a shape space based on the camber and thickness of the airfoil, where the cost of evaluating each candidate is the limiting factor. Compared with wind tunnel testing, high-fidelity three-dimensional CFD and panel methods coupled to an integral boundary layer, the hypothesis is that a consumer graphics card running a well-built two-dimensional simulator can perform the exploration stage by resolving the full flow field. There are five objectives: to develop a GPU-accelerated transient incompressible 2D CFD simulator able to evaluate an airfoil in minutes; to compute the aerodynamic forces by stress integration over embedded geometry; to couple the simulator to an evolutionary algorithm with physically motivated parameterisation and constraints; to obtain an optimised airfoil that outperforms its baseline at the design point; and to verify and validate the whole system. The methodology starts from Chorin's splitting on a stretched Cartesian grid domain, with semi-Lagrangian advection corrected by MacCormack, explicit diffusion with time-step subdivision and pressure projection solved by geometric multigrid. Turbulence is modelled with Spalart-Allmaras and the algebraic SA-BC transition criterion. The solid is handled by an immersed boundary method with ghost cells, and the forces are obtained by extrapolating the pressure to the wall, base pressure correction and integration of the viscous stress tensor. The optimisation is based on a genetic algorithm that uses a four-island model with sixteen individuals per island, with elitism, tournament selection, adaptive mutation step size and migration between islands every four generations, starting from four different airfoil families. The optimised airfoil then undergoes verification using a grid-convergence procedure based on Richardson extrapolation and the Grid Convergence Index (GCI) over four resolutions, with safeguards against pathological apparent orders. Validation shows that the computed lift converges across the test cases and agrees with the external reference within 5 %, whereas drag is overestimated by between 54 % and 146 % and does not converge in order, which limits conclusions to relative comparisons rather than absolute values. The optimisation campaign, run at Reynolds 100,000 and an angle of attack of 4 degrees, converges measurably: the shape distance between islands falls by a factor of fourteen over eleven epochs. The airfoil obtained is thin and strongly cambered, it is 2.27 times more efficient than its initial airfoil on the fine grid and 2.08 times in the extrapolated result, retaining its efficiency advantage across the nine angles evaluated, the four grids and the extrapolation. The total cost was 178 hours on a single consumer-grade GPU, confirming that full-field CFD shape exploration is feasible outside a dedicated computing infrastructure.

© CBLTIC Campus del Baix Llobregat - UPC