Analog and mixed-signal circuit design is notoriously difficult because performance depends on both the chosen device sizes and the circuit's internal electrical states. A new arXiv preprint introduces CktHGVAE, a heterogeneous graph variational autoencoder that explicitly models both kinds of variables. The authors argue that this symmetric design/state representation lets a single model handle both design and analysis tasks.
The model is aimed at transistor-level CMOS circuits, using a heterogeneous graph structure to represent the different types of components and their connections. By taking design variables and state variables as joint inputs and outputs, CktHGVAE differs from typical approaches that focus on design variables alone. The abstract does not include experimental details or comparisons, so the practical gains remain to be seen.
Because there is only one source, the article reflects that abstract's claims without independent verification. The main open question is whether the symmetric formulation actually improves accuracy or efficiency over existing circuit learning methods. Future work will need to show concrete results on real analog and mixed-signal benchmarks.