Automated Design of Two-Stage Op Amp Using Reinforcement Learning

In this work, a Reinforcement Learning-based method is proposed for automated parameter selection and optimization in the design of a two-stage operational amplifier. Our approach involves training an agent to interact with SPECTRE simulations by modifying netlists and simulation settings and interpreting output files. The respective design data inputs for our design engine, are only the topology and the performance specification space. Following training, the agent demonstrates superior performance compared to expert-driven approaches and generates a Pareto front of results for designers to select based on application requirements and trade-offs. This is achieved through a model that learns the intricate relationships between parameters and specifications. Our methodology is designed to operate independently from human intervention, offering a framework that is fully autonomous and converges to solutions in an average of 9 simulations, which is 3 × faster than similar approaches.