Analysis of Machine Learning Techniques for Time Domain Waveform Prediction in Analog and Mixed Signal Integrated Circuit Verification

Abstract

Pre-silicon analog and mixed signal (AMS) design verification involves exorbitant computing and manual effort and time to verify the design against the specification of an IC. This paper proposes a Machine Learning (ML) based behavioural model to predict the output response of AMS circuits that can be used in the automated verification process including automation of waveform review sign-off, and fast simulation models. The ML based behaviour model is constructed using the time domain features. To address both linear and non-linear behaviours of the circuit, this paper proposes a framework with statistical processing, waveform segmentation and circuit partitioning approaches as a divide and conquer strategy to identify the appropriate suite of ML algorithms. The best performing ML models in each segment are concatenated to stitch the complete response. We also propose SNR as a metric to evaluate the prediction accuracy. An Operational Amplifier (OpAmp) benchmark circuit has been used as a proof of concept to demonstrate this approach. An average SNR of 32 dB has been obtained in the prediction of the output waveform.