Prediction of parameters of semiconductor band-pass filters using artificial neural network

Abstract

Microwave device design process requires repetitive and time-consuming electromagnetic simulations to extract device parameters. New methods are needed to accelerate the process and enable it for real time parameter calculations. Feed-forward backpropagation multilayer perceptron artificial neural network with 3 hidden layers is presented to predict S21 parameters of 3.5 GHz band-pass filter. Prediction results, received with neural network trained with Resilient Backpropagation, One Step Secant, Scaled Conjugate Gradient and BFGS Quasi-Newton training methods, are compared with each other. The best prediction results are received after training network with BFGS Quasi-Newton training method. Average mean squared error and root mean square error received is 0.0044 and 0.066 respectively. As predictions of neural network are noticeably faster than electromagnetic simulation and prediction errors are low, the method can be used to predict band-pass filter parameters in real time.