Circuit2Graph: Diodes as Asymmetric Directional Nodes

Abstract

Circuit design requires both prototyping and simulation owing to the high degrees of freedom and mutual interference between the circuit components and wiring. In this paper, we propose a novel approach that involves transforming a diode into nodes within a graph network. This transformation considers the asymmetric directional characteristics of diodes in the forward and reverse directions. To evaluate the accuracy of our proposed method, we utilized graph classification with a graph neural network (GNN). Since a diode node holds two pieces of information, an anode and a cathode, the diode node is decomposed into two nodes, and the node attributes of the anode and cathode are assigned to each. By maintaining fixed hyperparameters and introducing a directional GNN and diode decomposition, the proposed method achieved an average accuracy of 98.80%. This accuracy surpasses the best accuracy reported in previous studies on graph classification by 0.9%. Notably, as with graph classification, changes in the handling of features obtained with GNN make it possible, for example, to optimize circuit topology and components and to predict the area required for implementation. Our method provides an environment for solving circuits on the GNN platform, enabling the prediction of values and states that could not be obtained with existing circuit simulations.

Keywords

• Asymmetric node
• directional graph neural networks
• directional node
• node separation
• series-parallel decomposition