EM-Based Parallel Neural Network Modeling Technique

Recently, people put forward higher requirement of the accuracy and the development cycle for RF and microwave electromagnetic(EM) modeling. Artificial neural network(ANN) has been widely used in RF and microwave field. ANN provides an accurate and fast modeling method for RF and microwave applications without understanding the internal details of the device. This technique has been proved to be a very effective modeling method for RF and microwave. However, ANN techniques for EM-based modeling usually need to repeatedly change the parameters of the microwave device and drive the EM simulator to obtain sufficient training and testing data. Nowadays, neural network technique for EM-based modeling is facing unprecedented challenges: on the one hand, with the growing scale of EM problem, this repetition of EM simulation takes a lot of time; on the other hand, the higher demands of modeling accuracy results in exponential growth of the training data which makes the neural network training process much slower.Although parallel computing technique has been widely deployed for high performance scientific applications, its application to RF and microwave is quite open. This paper aims at exploring the possibility of applying parallel computing technique to EM-based neural network modeling. Input samples and training samples are distributed to multiple processors under multiple computers for parallel processing through which the parallel EM data generation technique and parallel neural network training technique are implemented. Two microwave filter examples are used to verify our proposed EM-based parallel neural network modeling technique. The results show that our proposed EM-based parallel neural network modeling technique is practicable and can effectively improve the efficiency of EM-based neural network modeling while possessing the accuracy as same as the traditional neural network modeling technique.