Mixed-Variable Bayesian Optimization for Analog Circuit Sizing through Device Representation Learning

In this work, a deep representation learning method is proposed to build continuous-valued representations of individual integrated circuit (IC) devices. These representations are used to render mixed-variable analog circuit sizing problems as continuous ones and to apply a low-budget black box Bayesian optimization (BO) variant to solve them. By transforming the initial search spaces into continuous-valued ones, the BO's Gaussian process models (GPs), which typically operate on real-valued spaces, can be used to guide the optimization search towards the global optimum. The proposed Device Representation Learning approach involves using device simulation data and training a composite model of a Variational Autoencoder (VAE) and a dense Neural Network. The latent variables of the trained VAE model serve as the representations of the integrated device and replace the discrete-valued parametrizations of particular devices. A thorough explanation of the proposed methodology's mathematical formulation is given and example sizing applications on real-world analog circuits and integrated devices underline its efficiency.