Machine learning for true circuit EDA

This works combines neural ODEs and graph networks for circuit forward modeling. The purpose is to model circuits as graphs and learn to predict the evolution of a circuit state (currents, unkown voltages, as well as graph-level attributes such as power consumption). We first present the possible graph representations of a circuit that are explored in this work. We then present a modified interaction network (IN) framework that is used to learn the dynamics of circuits. In order to replicate the continuous nature of physical interactions occurring in circuits, we combine the IN framework with neural ODEs. We detail the combinations of implicit layers with an IN that we experimented. We list the experiments performed and describe their process. We finally analyse and compare the performances of our model.