Balancing Speed and Accuracy for Robust Analog-Mixed Signal Circuit Design using Closed-Loop Reinforcement Learning with Ensemble Neural Network Surrogates

Analog/mixed-signal (AMS) circuit sizing requires exploring high-dimensional, constrained design spaces under expensive SPICE simulations, while maintaining robustness to out-of-distribution (OOD) parameter proposals. We propose a fidelity-aware reinforcement learning framework (RL-PPO-Online) that unifies (i) SPICE-in-the-loop verification and (ii) ensemble neural network guided regularization for surrogate-assisted optimization. Another PPO policy (RL-PPO-Offline) is trained using an ensemble neural network for fast performance prediction, while a discriminator penalizes OOD actions to stabilize surrogate-only updates. We further introduce RL-PPO-Blend, which schedules a small fraction of SPICE queries (0.2-5%) on top of RL-PPO-Offline to provide a tunable speed-fidelity trade-off. On challenging ADC front-end sizing tasks in 28 nm (bootstrapped sampling switch and input buffer + switch), the proposed method improves pass rates over DDPG and LLM baselines, and achieves near-online accuracy with substantially reduced simulation cost.