Multi-band machine-learning framework for reliable 5–30 GHz LC-DCO synthesis in 22-nm FDSOI

Reliable millimeter-wave frequency synthesis requires oscillator designs that combine wide tuning coverage, low phase noise, and stable operation over broad frequency spans. This work introduces a machine-learning-driven framework for the automated design of LC Digitally Controlled Oscillators (DCOs) covering the 5–30 GHz range in 22-nm FDSOI technology. Surrogate models based on gradient-boosted ensembles are trained to predict oscillation frequency and phase noise directly from device dimensions, tank parameters, and bias conditions, enabling efficient navigation of the multi-dimensional design space. A targeted data-augmentation strategy enhances model generalization throughout the entire operating spectrum, while a frequency-aware decomposition further improves accuracy across the full range of synthesized oscillators. The proposed synthesis algorithm translates user-defined specifications, such as center frequency, tuning range, and phase-noise limits, into feasible transistor-level parameter sets and reconstructs the corresponding tuning boundaries via calibrated tank-capacitance adjustment. The automatically generated designs exhibit strong agreement with schematic and post-layout simulations, achieving wide tuning coverage and competitive phase noise with minimal deviation from predicted values. The results demonstrate that data-driven modeling supports reproducible, scalable, and specification-centric DCO design, offering a systematic alternative to conventional manual procedures and significantly reducing design effort in millimeter-wave oscillator development.