Regression Model-based VCO Design Optimization Technique

Along with technology advancements, implementing digital signal processing has become easier through digital circuit design automation. However, an analog circuit has too many obstacles to adopt a design automation flow in a scaled-down technology due to the relatively increased PVT (process-voltage-temperature) variations and enlarged mismatch errors, thus requiring a careful full-custom analog circuit design or employing an intensive back-end digital calibration engine. In this study, we implement an analog circuit design flow that quickly finds the optimal design point enabled by a high-level simulator programmed in Ocean scripts with a machine learning algorithm. To reduce the analog design effort, we train an ANN regression model that predicts the top-performing candidates with reasonable SPICE simulations in a supervised learning process. Here, we aimed to design a ring oscillator in the 65nm CMOS process. By using the proposed analog design automation strategy, we focused on jitter-power figure-of-merit (FOMJ) optimization, resulting in 99.84% accuracy of the final performance in comparison with SPICE simulation.