Modeling resonant frequency of microstrip antenna using GPU-based neural network

Resonant frequency is an important parameter in the design process of microstrip antenna (MSA). In order to deal with the issue of the long computing time and low accuracy of training neural network (NN) based on particle swarm optimization (PSO) algorithm when modeling the resonant frequency of rectangular MSA, parallel optimization based on graphic processing unit (GPU) is presented in this paper. The proposed method corresponds one particle to one thread, and deals with a large number of GPU threads in parallel to accelerate the convergence rate of the whole swarm and reduce the computing time of training NN. The resonant frequency of rectangular MSA is modeled based on the parallel PSO algorithm, and the experiments based on compute unified device architecture (CUDA) show that compared with CPU-based sequential PSO-NN, more than 300 times of speedup has achieved in GPU-based parallel PSO-NN with the same calculation precision. Substantially increasing the number of particles on GPU side can significantly reduce the network error with the very limited runtime increasement.