Machine learning modeling of power delivery networks with varying decoupling capacitors

Abstract

This paper presents modeling of power delivery network (PDN) impedance with varying decoupling capacitor placements using machine learning techniques. The use of multilayer perceptron artificial neural networks (ANN) and gaussian process regression (GPR) techniques are explored, and the effects of the hyperparameters such as the number of hidden neurons in the ANN, and the choice of kernel functions in the GPR are investigated. The best performing networks in each case are selected and compared in terms of accuracy using test data consisting of PDN impedance responses that were never encountered during training. Results show that the GPR models were significantly more accurate than the ANN models, with an average mean absolute error of 5.23 mΩ compared to 11.33 mΩ for the ANN.