Benchmarking the engineering reliability of LLM-based design for resonant wireless power transfer systems

Abstract

Designing high-frequency wireless power transfer (HFWPT) systems is a complex process that demands considerable time and domain expertise. To address these challenges, this study employs both general purpose and finetuned large language model (LLM) agents to automate the wireless power transfer (WPT) design process. The finetuned models embed WPT domain specific knowledge using instructions capable of generating complete coil and circuit level parameters including geometry, conductor type, compensation values, and frequency targets within seconds. Among the tested agents, high-frequency power electronics (HFPE-4o) achieved the best performance, reducing design time by a factor of 120 (representing a 99.17% reduction compared with conventional methods). The generated design achieved 99.98% accuracy, successfully producing all required parameters without additional input. Validation through power simulation (PSIM) and analysis system (ANSYS) simulations confirmed electromagnetic and circuit level accuracy, while 150 W experimental results verified system operation at 85 kHz with less than 8% deviation from predictions. Comparative evaluation against general purpose models such as ChatGPT-5 further demonstrated the advantage of domain adaptation, revealing significant improvements in numerical reliability and design stability. These findings confirm that LLM driven workflows can accelerate WPT system design while preserving engineering precision. By establishing a foundation for iterative refinement, the proposed artificial intelligence (AI)-driven agentic framework provides a promising pathway for integrating intelligent LLM agents into next generation power electronics design workflows.