In this thesis, the automated design of high-frequency digitally controlled oscillators (DCOs) with emphasis on phase noise optimization was implemented using a 22 nm Fully-Depleted Silicon-on-Insulator (FDSOI) CMOS technology process node. The study addresses cross-coupled LC voltage-controlled oscillators (VCOs) and DCOs incorporating a specialized capacitive tank designed to achieve an ex tended tuning range. A total of approximately 450 schematic-level VCO designs were generated by sweeping circuit parameters, including NMOS/PMOS transistor widths and lengths, inductor and capacitor values, bias current, and supply voltage. These designs covered oscillator frequencies from 5 GHz to 30 GHz and served as the dataset for training machine learning models. The thesis is organized into three major chapters. In the first chapter, the operating principles of cross-coupled LC VCOs are analyzed in detail. The chapter further examines the operation of the DCO capacitive tank, describing its contribution to tuning range and overall circuit behavior. This analysis provides a foundation for understanding the subsequent dataset generation and model training processes. In the second chapter, the development of machine learning models for frequency and phase noise prediction is presented. The frequency prediction model employs a stacking ensemble of XGBoost and LightGBM base learners, with meta and residual XGBoost models, while the phase noise model is implemented using XGBoost. Hyperparameter tuning was performed using Optuna, and model vali dation was conducted with K-fold cross-validation. A frequency generator module was developed to perform inverse design using Bayesian optimization (TPE sampler). The first DCO, with a tuning range of 19–30 GHz, was used to validate the predictive models. Subsequently, two additional DCOs, covering tuning ranges of 11–19 GHz and 5–11 GHz, were generated by the final ML framework, demonstrating its capability for automated design across different frequency bands. In the third chapter, the generated DCO designs underwent full layout implementation, parasitic extraction, and post-layout simulations. Electromagnetic analysis using Raptor X was conducted in three scenarios: inductor only, capacitor only, and combined. This approach allowed evaluation of the contributions of each element to circuit performance. Layout improvements were applied to enhance electromagnetic behavior. The validation results confirm that the ML-assisted frame work successfully produces DCO designs that satisfy the target frequency and phase noise specifications, demonstrating the effectiveness of machine learning-based automated oscillator design.
Παναγιώτα Κ. Τσίμπου (Wed,) studied this question.