Next Generation of AI for Power Electronics: Explainable, Light, and Flexible

The presentation titled "Next Generation of AI for Power Electronics: Explainable, Light, and Flexible" addresses the challenges of existing AI integration in power electronics, particularly issues related to non-explainability, heavy data and computation dependency, and inflexibility. To overcome these, Physics-Informed Machine Learnings (PIMLs) are well-positioned as innovative solutions, embedding physics-based knowledge directly into AI models and serving as next generation of AI for power electronics fields. PIML approaches are classified into three main types: Physics-in-Loss (integrating physical laws in the loss function), Physics-in-Initialization (using physics-based initialization), and Physics-in-Architecture (designing neural network structures aligned with physical principles). Implications, applications, and pros and cons of these PIML methods are discussed. Considering the nature of power electronics domains with rich physics foundations but lacking abundant data support, the Physics-in-Architecture methods lie at the center of this presentation, which progressively discusses its power-electronics-specific explainability, data-light and lightweight advantages, and training-free zero-shot flexibility across diverse scenarios. Potential future directions for applying PIMLs in the lifecycle of power converters are shed light on, paving the way for democratizing next-generation AI in power electronics. Please note, the code demos provided are open sourced in the GitHub repositories: https://github.com/XinzeLee/PANN https://github.com/XinzeLee/PE-GPT