Few-shot learning GNN-EQL model with gm/ID method for analog integrated circuit design

Abstract

Analog integrated circuit design typically involves extensive analytical derivations to evaluate circuit performance metrics. Although SPICE simulations facilitate efficient prediction of these metrics, the simulation process remains time-consuming as the dimensionality of circuit parameters increases. In this article, we propose integrating Graph Neural Networks (GNN) with Equation Learner Networks (EQL), employing them as pretrained models within a limited range of design parameters. Our results demonstrate that datasets constructed using the $g_m/I_D$ method capture design points more efficiently compared to the random sampling of width-to-length ($W/L$) ratios. Furthermore, experimental validation indicates that the pretrained GNN-EQL model achieves superior performance compared to other pretrained models when the parameter range expands across three different amplifier designs. Finally, our approach significantly reduces the required samples by up to 20X when adapting the pretrained model to broader parameter ranges, compared to training a new model from scratch.