Improvement of Neural Networks Applied to Photoacoustic Signals of Semiconductors with Added Noise

Abstract

This paper provides an overview of the characteristics of different neural networks trained on the same theoretical database of ntype silicon photoacoustic signals. By adding different levels of random Gaussian noise to the training input signals, two important goals were achieved. First, the optimal level of noise was found which significantly shortens the training networks with minimal loss of accuracy of its predictions. Second, the termination criteria of networks training were activated to avoid overtraining, i.e., networks generalization was performed. A networks efficiency analysis was performed on both theoretical and experimental photoacoustic signals, resulting in a selection of one neural network that is optimal to the performance requirements of the real experiment. It is indicated that the application of such trained networks is more reliable on thicker semiconductors, whose thickness is greater than the value of the carrier diffusion length in the investigated sample.