RESEARCH / SEMICONDUCTOR
How complete circuit-optimization episodes are collected and checked against a simulator, creating traceable training examples that preserve proposals, measurements, corrections and the evidence behind each design target.
Read on this pageTraining an AI circuit-design agent requires more than a collection of finished designs. It needs records of what was proposed, what the simulator measured, why a change was tried and how the design recovered when it failed. This paper describes how that evidence is collected for analog circuit optimisation.
The collection spans complete optimisation episodes on SKY130 circuits, preserving measurements, target-level scores and the model's hypotheses alongside each action. A two-round process distinguishes original design ambitions from targets that a given circuit has actually demonstrated it can reach. The result is an account of how training data can remain auditable: readers can follow a design trajectory and inspect the basis for its reported progress rather than relying on a final score alone.
Complete episodes preserve proposals, simulator results, corrections and termination reasons.
Scores retain the contribution of each design target so progress can be checked.
Final targets are recorded with their provenance and kept distinct from original specifications.
The corpus and measured outcomes concern simulator runs on the documented SKY130 circuit collection. Finalised targets can differ from the original datasheet targets; convergence against those targets is not proof that every original specification was met or that a design has been validated in silicon.
The complete paper, with its original figures, tables and references.
38 pages · Read without leaving this pageDiscuss the approach, evidence and relevance to your team with the people building it.
