SourceCircuit Representation Learning with Masked Gate Modeling and Verilog-AIG Alignment
The paper evaluates DeepGate2's ability to identify logically equivalent gate pairs by computing cosine similarity between their AIG embeddings. On 10000 pairs of logic gates derived from 10 circuit designs, DeepGate2 achieves an average precision of 0.295, recall of 0.841, F1-score of 0.424, and AUC of 0.804. Performance varies widely across designs, with i10 being the strongest (F1=0.575, AUC=0.918) and apex1 the weakest (F1=0.223, AUC=0.601).