This fully numerical approach could pave the way for compact, low-cost, and widely adopted wireless charging solutions within the next decade.
A research team from Chiba University has developed a machine learning-based method to design more efficient and stable wireless power transfer (WPT) systems. Traditional WPT systems rely on complex analytical equations to maintain stable output voltage and achieve zero-voltage switching (ZVS) under changing loads. However, these equations often assume ideal conditions and fail to capture real-world complexities, making load-independent (LI) operation difficult to achieve. To overcome this, the researchers proposed a fully numerical design approach that uses differential equations to model real circuit behavior, including parasitic effects, and applies a genetic algorithm to optimize key performance metrics like voltage stability, power efficiency, and harmonic distortion.
The method was tested on a class-EF WPT system, combining a class-EF inverter with a class-D rectifier. Conventional class-EF systems typically maintain ZVS only at their rated load; any deviation leads to a drop in efficiency and unstable output voltage. In contrast, the LI system designed through this machine learning-based process maintained ZVS and kept voltage fluctuations under 5%, compared to 18% in traditional setups. It also demonstrated consistent performance across varying loads by keeping the output current steady, resulting in a power-delivery efficiency of 86.7% at 6.78 MHz and delivering over 23 watts of output power.
The numerical design approach not only simplifies circuit development by removing the need for fine-tuning but also enables more compact and cost-effective systems. By accurately capturing the impact of diode parasitic capacitance, it performs better even under light loads. Professor Sekiya emphasized that this breakthrough marks the first successful fully numerical design using machine learning in power electronics. He believes load-independent operation is essential for real-world WPT deployment and envisions widespread adoption within the next 5 to 10 years. More broadly, the study highlights the transformative potential of AI and machine learning in automating power electronics design—ushering in a future where smart, self-optimizing systems become the norm.
