Robust Automated Synthesis for Millimeter-Wave VCOs Using Hidden-Constrained Machine Learning-Assisted Optimization

The automated synthesis of millimeter-wave voltagecontrolled oscillators (VCOs) faces a severe efficiency bottleneck caused by oscillation startup failures. These failures constitute a "hidden constraint" as non-oscillating designs fail to produce valid performance metrics, rendering standard optimization feedback ineffective. To address this, we propose a hiddenconstrained machine learning-assisted optimization (HC-MLAO) method for robust automated VCO synthesis. This framework utilizes a composite radial basis function model to characterize the hidden startup boundary, transforming it into an explicit, optimizable condition. Furthermore, we introduce improved acquisition functions incorporating startup probability, alongside a batched design search strategy to efficiently target the feasible space and enable parallel verification. The effectiveness and generality of the proposed method are demonstrated through two custom benchmark functions and three practical VCOs with class-C, class-F, and multi-core topologies. Post-layout validation confirms that the HC-MLAO method delivers robust solutions with significantly enhanced optimization efficiency and superior circuit performance compared to conventional methods.