Designer-Inspired AI-Assisting Methods for Power-Efficient RF Oscillator Design

To address the limitations of traditional manual tuning, mode-specific design, and fixed-performance targeting in analog/RF circuit design — such as restricted output power, discontinuous efficiency profiles, and slow convergence in existing ultra-low-power 2.4GHz BLE transmitter designs — this talk proposes a designer-inspired reinforcement learning framework centered on a waveform-biased graph neural network for the automated design of a power voltage-controlled oscillator (PVCO). Unlike prior machine-learning-based circuit optimization approaches that lack waveform awareness and cross-configuration adaptability, the designer-inspired framework models PVCOs as graph structures (ie devices as nodes, RF/physical/GND interconnects as edges) to preserve topological connectivity, while integrating waveform insights into two critical stages: 1) the policy network for action generation, and 2) reward shaping to complement conventional designer insights and guide learning. This mimics the decision-making of experienced circuit designers while retaining AI-driven automation.