Open Access

Monish Suresh -

• In this bachelor thesis, I investigate the feasibility of Gaussian Process Regression (GPR) (Williams and Rasmussen 2006) in the control of a buck converter (R. W. Erickson and aksimovic 2020). I develop a prototype digital controller for a buck converter, which uses a trained GPR model to predict the PWM duty cycle required to obtain a desired output voltage. I use an Analog Discovery 2 (AD2) (Digilent Inc. n.d.) as an interface between the GPR controller and the buck converter hardware by providing real-time voltage measurements to the controller and generating the PWM signal. I compare the performance of the GPR controller with the TPS629210DRLR Integrated Circuit (IC) (Texas Instruments 2022), which uses an analog control method and serves as the reference performance standard. I compare the performance of the GPR controller and the IC based on their ability to maintain a target voltage during a load change involving an RC load. The results are evaluated based on the voltage regulation accuracy and load voltage settling time. The results show that both controllers are able to set the load voltage within 1% of the target value. However, the GPR controller has a settling time around 3 times longer than the analog controller. The longer settling time is due to latencies from communication with the AD2 and the time required for duty cycle computation. Despite this, the settling time difference is short enough to consider the digital GPR controller competitive to the reference analog control method, especially since the current prototype is not optimized to reduce these latencies. The findings of this thesis highlight both the potential and limitations of the digital control of a buck converter.