LSTM-RNN Based Analog IC Automated Sizing Model for Operational Amplifier and VCO

Abstract

Artificial intelligence and machine learning have been widely used to replace human work. Analog integrated circuit design needs to adjust a large number of circuit parameters to satisfy the balance between various performance metrics, which now mostly rely on the experience of designers and sometimes of the intuitions. Machine learning has demonstrated the potential to aid the design of analog integrated circuits in the literature, whilst most of them adopted particle swarm intelligence and Bayesian optimization. However, the model training time and simulation run time are substantial when any circuit structure modification occurs. In this paper, Recurrent Neural Network (RNN) was used to automatically optimize the parameters sizing by giving requested performance. Training data sets for the RNN of component parameters and circuit performance were simulated using Cadence Spectre. After training for only 15 minutes, RNN learns to predict parameters by inputting gain, bandwidth, power and frequency, which can make critical circuit design decision significantly faster. The reliability and applicability of the algorithm was verified through the parameter prediction of integrated operational amplifier and VCO.