Hybrid Utilization of Subgraph Isomorphism and Relational Graph Convolutional Networks for Analog Functional Grouping Annotation

A hybrid method is proposed that combines the subgraph isomorphism algorithm VF2 and a relational graph convolutional network (RGCN) model for the recognition and classification of functional pairs in an analog circuit at the device level. To apply the RGCN model, a heterogeneous graph representation is proposed that includes ten edge types. The RGCN model is trained to predict the presence or absence of a functional pairing between two transistors. With the proposed hybrid approach, the RGCN model is utilized to filter and label the functional pairs returned by VF2. The techniques are characterized on a dataset of 14 circuits that include a total of 219 transistors and 120 functional pairs that are manually annotated. While VF2 achieves a perfect recall of 1, recognizing all 120 functional pairs, an average false positive rate of 35.9% is observed among the detected pairs. The hybrid approach that integrates VF2 with the RGCN model results in an F1 score of 0.882, a 14.4% improvement over VF2 when executed alone. In addition, the hybrid approach returns the specific functional type of each detected pair. The average execution time of the hybrid approach is 0.594 seconds. The results confirm the effectiveness of the proposed hybrid approach in detecting analog functional pairs in practical analog EDA applications including for annotation of symmetry constraints. The detected and labeled functional pair types also provide utility as features for learning models utilized in additional circuit applications.