Physics-Informed Autonomous LLM Agents for Explainable Power Electronics Modulation Design

TL;DR

PHIA tackles explainability and data efficiency in AI-driven power electronics design by coupling an LLM-based planner with two hierarchical physics-informed neural networks for modulation design of the DAB converter. Two PINNs, ModNet and CirNet, model switch-level dynamics and circuit-level physics and feed an optimization loop to output modulation parameters with transparent explanations. Empirical results show a MAE reduction of $63.2\%$ in low-data scenarios and a design-time reduction of more than $33\times$, validated on hardware-based TPS modulation data. A 20-expert user study confirms substantial efficiency gains and usability, indicating PHIA's potential to transform industrial PES design workflows.

Abstract

LLM-based autonomous agents have recently shown strong capabilities in solving complex industrial design tasks. However, in domains aiming for carbon neutrality and high-performance renewable energy systems, current AI-assisted design automation methods face critical challenges in explainability, scalability, and practical usability. To address these limitations, we introduce PHIA (Physics-Informed Autonomous Agent), an LLM-driven system that automates modulation design for power converters in Power Electronics Systems with minimal human intervention. In contrast to traditional pipeline-based methods, PHIA incorporates an LLM-based planning module that interactively acquires and verifies design requirements via a user-friendly chat interface. This planner collaborates with physics-informed simulation and optimization components to autonomously generate and iteratively refine modulation designs. The interactive interface also supports interpretability by providing textual explanations and visual outputs throughout the design process. Experimental results show that PHIA reduces standard mean absolute error by 63.2% compared to the second-best benchmark and accelerates the overall design process by over 33 times. A user study involving 20 domain experts further confirms PHIA's superior design efficiency and usability, highlighting its potential to transform industrial design workflows in power electronics.

Figures (5)

• Figure 1: An example of power converter application: the DAB (Dual Active Bridge) converter serves as a DC transformer of the DC-DC power grid, offering galvanic isolation and regulating power and voltage between DC buses. Modulating its switches directly affects the system operating performance, including power transfer efficiency, voltage regulation, and stability of the interconnected buses.
• Figure 2: System architecture of PHIA: an engineer provides design requirements to PHIA via a chat interface connecting to its planner. Once the full requirements are determined, the planner coordinates and invokes tools from the tool set to iteratively generate the modulation design without human supervision. After the design is done, the planner displays the final results and explainable process on the chat interface.
• Figure 3: The proposed surrogate model consists of two physics-informed neural networks, namely, ModNet for switch-level modeling to learn the switching behaviors, and CirNet for system-level modeling to learn the circuit physics. The hierarchical structure enhances the overall accuracy of the power converter's modeling of the complex behaviors of the switches.
• Figure 4: PHIA with different structures of CirNet. The minimal MAE of 0.235 on validation set is reached when there are 2 hidden layers with 32 hidden neurons.
• Figure 5: Modeling performance of the top 5 benchmarks and the proposed PHIA. The modelling results of PHIA is closest to the measurement, demonstrating its outstanding modeling accuracy.