AI-Driven RFIC and Electromagnetic Design Framework Adapted to an Open-Source PDK

Abstract

The design of RFICs is traditionally a labor-intensive process that relies on human intuition and iterative manual tuning using commercial Electronic Design Automation (EDA) tools. Recent work has shown that Reinforcement Learning (RL) has the potential to automate portions of this process, but existing implementations rely heavily on proprietary Process Design Kits (PDKs) and licensed simulation software such as Cadence Spectre, limiting accessibility and reproducibility. Building on an existing workflow developed by Zhou et al. [1], this thesis presents a fully open- source RL-based optimization framework for active circuit optimization in RFIC design, combining the open-source IHP 130nm SiGe:C BiCMOS PDK with the open-source simulator Ngspice. A Proximal Policy Optimization (PPO) agent is trained to perform multi-objective optimization of single- and two-stage common-emitter Low-Noise Amplifiers, simultaneously maximizing S21 (forward gain) at a target frequency, minimizing S11 and S22 (input and output return loss), minimizing noise figure, and reducing DC power consumption through a weighted reward function. Experimental results demonstrate that the agent can converge from arbitrary, detuned initial configurations to near-optimal designs. The work validates that RL-driven circuit optimization is feasible within a fully open-source toolchain and provides a reproducible foundation for further research.