Reproduce “DiffCkt: A Diffusion Model-Based Hybrid Neural Network Framework for Automatic Transistor-Level Generation of Analog Circuits” on OpenResearch

About this paper

Chengjie Liu, Jiajia Li, Yabing Feng, Wenhao Huang, Weiyu Chen, Yuan Du, Jun Yang, Li Du

DiffCkt generates amplifier topologies and transistor sizes from specifications. It predicts component count, then uses discrete graph diffusion for connections and continuous graph diffusion for device parameters. For 50 requests at DDIM interval step 5, it reports a relative-performance score of 0.804 ± 0.181 and generation time of 1.57 ± 0.08 seconds for simulated TSMC 65-nm operational amplifiers, without guaranteeing valid circuits.

Key insights

• Treating transistor-level analog circuit synthesis as an inverse problem allows diffusion models to effectively learn complex relationships between circuit performance, structure, and parameters, enabling direct generation from specifications.
• Decomposing the overall generation task into distinct sequential stages—device number prediction, discrete graph structure generation, and continuous parameter sizing—simplifies model training and enhances the tractability of the complex problem.
• The developed graph representation, particularly the use of port-specific edge matrices, is crucial for capturing the unique and intricate connectivity of analog circuits, which is more detailed than typical graph structures.

Summary by alphaXiv.