RESEARCH / SEMICONDUCTOR
How circuit-design traces train a specialist model, with simulator feedback guiding improvement and a larger model joining when progress stalls. The paper examines the hybrid system’s engineering results.
Read on this pageAnalog circuit optimisation requires repeated proposals, simulation and correction. This paper examines how those design traces can train a specialist model, and how that model can work with a larger model when progress stalls rather than relying on the larger model for every step.
The method combines supervised learning from traces with reinforcement learning against circuit-simulator feedback. It also studies a reward-design problem: a failed measurement must not look better than a genuine but imperfect result. The deployment strategy preserves the specialist's best state when escalating a difficult circuit. A controlled ten-circuit comparison explores the resulting trade-off between meeting design targets and using an external model, providing an engineering case study of specialist AI with a defined fallback.
Simulator feedback ties model training to measured circuit behaviour.
Reward checks prevent missing measurements from being mistaken for design improvements.
A larger model joins only when the specialist's recorded progress shows it is needed.
The reported comparison covers ten circuits under the paper's simulator, targets and iteration budgets. Results describe the hybrid system, not the specialist acting alone, and do not establish silicon performance or generalise automatically to other process technologies and circuit families.
The complete paper, with its original figures, tables and references.
28 pages · Read without leaving this pageDiscuss the approach, evidence and relevance to your team with the people building it.
