Machine Learning based Waveform Predictions using Discrete Wavelet Transform for Automated Verification of Analog and Mixed Signal Integrated Circuits

Present day verification process of Analog and Mixed Signal (AMS) ICs involve significant manual effort and designer's expertise. One of the crucial processes in automating verification of AMS ICs is waveform prediction. Waveforms of ANIS circuits have time varying amplitude and frequency information. Hence a promising tool for extracting the information in a compressed dataset is a Discrete Wavelet Transform (DWT) representation. The present work performs waveform prediction by extracting DWT coefficients from the time domain waveform of an AMS circuit and uses a reduced dataset to derive a prediction model using machine learning (ML). An OpAmp and Comparator from analog benchmark circuit are used in this present work as proof of concept using ML approaches such as Random Forest and RNN. ML model is trained with input and output waveforms of functional simulations across process corners, supply voltage and temperature variations. In this proposed work, the suggested approach demonstrates a highly accurate resemblance between the predicted waveform and the real waveform. The best case prediction yields an RMSE of 0.01, an R-2 Score of 1, and a signal-to-noise ratio (SNR) of 65 dB. Compared to other methods reported in literature that use raw transient analysis simulation data, the method proposed in this paper results in higher prediction accuracy by 39 dB of SNR.