A Hybrid Deep-Belief and Knowledge-Based Neural Network for Efficient Prediction of Jitter in the Presence of Multiple PDN Noise Sources

Abstract

In this article, an efficient approach is developed to predict the jitter in the presence of multiple noise sources, such as power supply noise, ground bounce noise as well as input data noise in diverse power delivery modules by combining the knowledge-based neural network with the deep belief neural network. The proposed hybrid neural network achieves reasonable accuracy while providing for efficient training using input data obtained from both analytical closed-form expressions as well as a circuit simulator. The proposed model can also handle varying inputs without retraining the network's parameters. In order to optimize the training dataset, instead of using the random dataset, a new configuration with a mixed dataset (with a combination of uniformly distributed data as well as randomly distributed data) is proposed. Their performance along with different types of energy models is also investigated.