Homogeneous Versus Heterogeneous Graph Representation for Graph Neural Network Tasks on Electric Circuits

Machine learning (ML) has the potential to revolutionize electronic design automation (EDA), but mainly lacks the scalable and adequate representation of electric circuits, and requires diverse datasets that capture the performance of various electrical circuits. Literature shows methodologies that fall short on representing unlimited complexity circuits, in which models are hard coded to a limited certain circuit order, and are more focused on using a fixed circuit topology rather than providing a general purpose representation. This paper proposes a generalizable and scalable circuit graph representation based on bond graphs modeling as heterogeneous graph. Additionally, an analytical comparison between mapping the bond graph as either homogeneous or heterogeneous graph is shown. By analyzing the graph in spectral domain, the frequency component distribution of node features is found to impact the performance of Graph Neural Networks (GNNs) unique to the graph representation used. The analysis discusses inherent limitations of GNNs, and reveals that GNNs designed for heterogeneous graphs can more effectively capture system dynamics & variations. Finally, different case studies are presented utilizing heterogeneous graph neural network (Hetero-GNN) for different ML tasks including dynamics prediction and converter classification.