Researchers at Princeton have developed AI-driven methods to automate RF integrated circuit (RFIC) design, a field long considered a 'dark art' requiring years of human expertise. Using reinforcement learning to determine circuit architecture and topology, combined with a convolutional neural network-based electromagnetic emulator for inverse design of passive structures, the system can generate fabrication-ready RFIC layouts from specifications in minutes rather than months. A 2023 proof-of-concept 5G power amplifier achieved record bandwidth-efficiency performance with unconventional, non-human-like layouts. A diffusion model was later added to allow designers to control the interpretability of generated structures. Key challenges remain: AI hallucinations require human verification, and progress toward a universal foundational model is bottlenecked by simulation data locked behind NDAs. The authors call for open data ecosystems in the RFIC community.
