Using Denoising Diffusion Probabilistic Models to solve the inverse sizing problem of analog Integrated Circuits

Abstract

In this paper, we focus on using Artificial Neural Networks (ANNs), particularly diffusion models, to automate the sizing of analog Integrated Circuits (ICs), given the constraints of their performance metrics. Researchers have explored various automation methods, including meta-heuristics and optimization-based approaches, to address this challenge. However, each method presents distinct drawbacks and, at times, yields inefficient results. While studies have made some attempts using ANNs, they commonly face the hurdle of the ill-posed nature of the problem exacerbated by the scarcity of databases for training the models. Therefore, this work introduces a novel approach to automate the design process by leveraging diffusion models to enhance the existing ANNs-based framework and address the limitations of previous methodologies. Specifically, Denoising Diffusion Probabilistic Model (DDPM) to tackle the inverse problem of the analog IC sizing. DDPMs employ a noising and denoising architecture, where they learn to reconstruct input distributions by progressively adding and removing noise. Once trained, the DDPM can generate new data from pure noise. We show that even a simple DDPM can sample sizing solutions in a small amount of time, which is an important steppingstone for future research. Experimental results indicate that our models successfully sized the two tested circuits with an average median error of around 6%, surpassing the state-of-the-art approaches whose error was over 60 to 70% higher. Moreover, by taking advantage of the generative capabilities of our models, we were able to generate points for targets within the dataset, with most of them showing an error below 3%. For the more challenging targets, we managed to find solutions with errors below 10%, while the supervised approaches struggled to achieve errors under 20%.