Knowledge-Based Bidirectional Recurrent Neural Network Approach for Efficient Prediction of Jitter in a Chain of CMOS Inverters

Abstract

An efficient hybrid approach based on combining the bidirectional recurrent neural network with knowledge-based neural network is presented to predict jitter in a chain of CMOS inverters in the presence of multiple noise sources. The new method achieves a reasonable accuracy and provides for efficient training using input data obtained from both a circuit simulator as well as analytical relations. The proposed approach can also estimate jitter for each inverter in the chain by only employing the accurate training data associated with the first inverter, resulting in a significant increase in speed compared to conventional approaches.