CFD-based Aeroelastic Analysis Based On System Identification And GPU Parallel Computing

This paper starts with flutter solution of transonic. By comparing with several methods, CFDmethod is chosen to solve the unsteady aerodynamic. To solve the problem that CFD iscomputationally intensive and inefficient, two methods of both realization and theoretically are used.Firstly, on realization, this paper uses GPU parallel computing based on CUDA/C programminglanguage instead of traditional CPU serial computing. By comparing these two methods, we find thatGPU parallel computing is faster and more efficient than traditional CPU serial computing, and it laysthe foundation for the deeper research. Secondly, ROM based on system identification is used toreduce the order of CFD solver by finding the relationship between the input and output. Twomethods are introduced: ARMA model and Kriging surrogate model. This paper compared theadvantage and disadvantage of these two methods. ARMA model is easy to get, and the accuracy andefficiency are both good. Kriging surrogate model goes a further step than ARMA model that it canget the ROM of multi-mach number and be more efficient. One can get higher speed-up by combiningGPU parallel computing and ROM.To verify the correctness of the theory above, NACA0012airfoil and NACA64010airfoil arechosen to check the unsteady fluid solver and the flutter response and flutter boundary. The result ofthis paper agrees well with the result of experience and literatures.