ZeroSim: Zero-Shot Analog Circuit Evaluation with Unified Transformer Embeddings

The introduction of ZeroSim marks a significant advancement in the field of analog circuit design automation, particularly in performance evaluation, which has been a major bottleneck due to the time-consuming nature of traditional SPICE simulations. ZeroSim utilizes a transformer-based framework that allows for robust in-distribution generalization across trained topologies and zero-shot generalization to unseen topologies without the need for fine-tuning. This is achieved through a comprehensive training corpus of 3.6 million instances covering over 60 amplifier topologies, as well as innovative strategies such as unified topology embeddings and topology-conditioned parameter mapping. The framework's ability to deliver accurate predictions across different amplifier topologies positions it as a superior alternative to existing machine learning methods, which often require extensive retraining or manual adjustments. By significantly outperforming baseline models, ZeroSim not only enha…