PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Junhua Liu, Fanfan Lin, Xinze Li, Kwan Hui Lim, Shuai Zhao

arXiv:2411.14214v2cs.AIcs.ET

TL;DR

Modulation-design automation for power electronics faces challenges in explainability, scalability, and practical usability as renewable-energy systems grow more complex. PHIA combines an LLM planner with physics-informed simulation and optimization; it reports lower error than baselines and faster design in a 20-expert study.

  • Problem

    Existing AI-assisted modulation-design methods face challenges in explainability, scalability, and practical usability, while growing and interconnected renewable-energy systems make manual design increasingly complex.

    Source: Introduction · p. 1

  • Method

    PHIA uses an LLM planner to process user requirements and coordinate physics-informed surrogate models and optimization tools that iteratively refine modulation parameters.

    Source: Introduction · p. 2, Physics-Informed Power Electronics… · p. 6

  • Results

    PHIA reports 63.2% lower error in low-data scenarios and 23.7% lower error in high-data scenarios than the best baseline; a study with 20 experts found design time over 33 times faster than conventional methods.

    Source: Introduction · p. 2

  • Takeaways & Limitations

    PHIA returns design parameters and performance metrics with visual explanations, and the paper reports efficient design with minimal human supervision.

    Source: Introduction · p. 2

Abstract

from arXiv · show

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.

Introduction

The introduction frames modulation design as increasingly difficult amid growing renewable-energy systems and identifies limitations in existing AI approaches. It presents PHIA as an LLM-based agent that coordinates physics-informed models and optimization to produce explainable designs with minimal human supervision.

Introduction

  • Growing renewable-energy adoption and interconnected converter systems increase modulation-design complexity, making manual optimization infeasible and requiring designs tailored to potentially conflicting application objectives.Power Electronics Systems support renewable integration, energy-storage management, and grid stability through converter modulation.Source: Introduction · p. 1
  • Existing AI modulation methods face data-intensive training, computational and energy costs, limited explainability and scalability, and substantial human involvement.The introduction cites LLM-based design, XGBoost with differential evolution, and offline Q-learning as examples of prior approaches.Source: Introduction · pp. 1–2
  • PHIA uses an LLM planner to turn user requirements into design specifications, then iteratively coordinates physics-informed surrogate models and optimization tools and presents parameters and performance metrics.The displayed results and process are intended to make the design process explainable.Source: Introduction · p. 2
  • PHIA achieves 63.2% lower error in low-data scenarios and 23.7% lower error in high-data scenarios than the best baseline; a 20-expert study reports design time over 33 times faster than conventional methods.The paper also reports strong performance for its hierarchical surrogate model in sparse-data and extreme scenarios.Source: Introduction · p. 2

Problem Statements

The section describes phase-shift modulation for DAB converters and formulates the design task around operating conditions, a selected strategy, and prioritized performance objectives. It also specifies desired features for an automated design system.

Power Converter Phase Shift Modulation

  • DAB phase-shift modulation adjusts switch timing to regulate current flow and power transmission; strategies differ in degrees of freedom and modifiable phase parameters.The DAB connects DC buses, while its operation affects power-transfer efficiency, voltage regulation, and bus stability.Source: Power Converter Phase Shift Modulation · p. 2, Figure 1 · p. 2, Table 1 · p. 3
  • Design inputs include rated and actual operating conditions, a modulation strategy selected for control complexity and achievable performance, and prioritized objectives such as efficiency, switching behavior, stress, and stability.Other listed objectives include power density, dynamic response, electromagnetic interference, and compatibility.Source: Power Converter Phase Shift Modulation · p. 2

Problem Statements

  • PHIA’s architecture takes requirements through a chat-connected planner, which invokes tools iteratively and returns the final design and an explainable process.The system is intended to generate the modulation design without human supervision after requirements are determined.Source: Figure 2 · p. 3
  • The design problem is to choose modulation parameters for the selected strategy that optimize prioritized objectives under specified converter operating conditions.The stated conditions include rated and actual power and input and output voltages.Source: Problem Statements · p. 3
  • The desired system should generate precise outcomes automatically, remain usable and explainable, and require little setup or extensive data.These features are intended to reduce complex user computation and support engineering design and scientific discovery.Source: Problem Statements · p. 3

Related work

The related-work section reviews LLM-based agents that plan and invoke tools for complex tasks, alongside physics-informed neural networks that incorporate physical knowledge into learning. It situates PHIA’s hierarchical surrogate model against these research directions.

LLM-based Autonomous Agents

  • LLM-based agents extend language models from reasoning and content generation to planning and tool use, with prior work applying them to industrial and domain-specific tasks.Examples include robot task generation, interactive decision-making, spatial-temporal planning, customer support, and financial trading.Source: LLM-based Autonomous Agents · pp. 3–4
  • Prior research describes LLM agents as combining planning with components such as memory and tools, with retrieval-augmented generation among techniques used for complex tasks with little instruction.The section also surveys LLM work in programming, multilingual tasks, and mathematical problem-solving.Source: LLM-based Autonomous Agents · p. 3
  • PHIA’s surrogate model uses hierarchical physics-informed networks: ModNet learns switch-level behavior and CirNet learns system-level circuit physics.The cited figure states that this hierarchy enhances modeling accuracy for complex switch behavior.Source: Figure 3 · p. 4

Physics-Informed Neural Networks

  • PINNs incorporate physical principles through components such as loss functions, parameter initialization, or network architecture, enabling learning of features governed by physical laws.The related work notes that conventional neural networks are commonly data-hungry and not explainable.Source: Physics-Informed Neural Networks · p. 4
  • Prior PINN studies address convergence, stability, and accuracy and span applications including traffic, power systems, lakes, reservoirs, quantum circuits, and turbulent flows.The cited work includes circuit-parameter determination for power-converter health monitoring and power-flow computation in electrical systems.Source: Physics-Informed Neural Networks · p. 4

Methodology

PHIA links an LLM planner to optimization tools and physics-informed surrogate models to generate modulation designs. Its hierarchical models predict converter behavior, while the system evaluates performance and presents design outputs to users.

Agentic AI System

  • PHIA collects user specifications, uses GPT-4 for planning, and invokes surrogate models and optimization tools to return and visualize modulation parameters.The interface displays charts for explainability and readability.Source: Agentic AI System · p. 4

Physics-Informed Power Electronics Modeling

  • The hierarchical surrogate pairs ModNet, which learns switch-level behavior, with CirNet, which models circuit physics and infers key state waveforms.Those waveforms support evaluation of current stress, soft-switching range, and efficiency.Source: Physics-Informed Power Electronics… · p. 5
  • ModNet uses LN-GRU layers to predict primary and secondary bridge voltages, combining physics-based synchronization loss with oscilloscope data to capture practical oscillations and overshoots.Its intermediate predictions within a switching cycle are stored for training.Source: Physics-Informed Power Electronics… · p. 5
  • CirNet takes ModNet outputs and iteratively predicts subsequent states using circuit dynamics embedded in physics-informed training losses.Training also uses experimental inductor-current points as ground truth.Source: Physics-Informed Power Electronics… · p. 5
  • Optimization algorithms iteratively pass operating conditions and modulation parameters to ModNet and CirNet, using evaluated metrics to steer the search.The system evaluates current stress, conduction loss, and soft-switching performance; adding surrogate models for further functions is described as scalable without changing the user experience.Source: Physics-Informed Power Electronics… · p. 6

Experiments Setup

The experiments assess PHIA on DAB-converter inductor-current forecasting using hardware waveform data and compare it with established time-series and machine-learning baselines. Models are tuned on validation data, and waveform plots compare predictions with measurements.

Data preparation

  • The forecasting task uses 200 hardware-acquired DAB-converter waveform sequences under triple phase-shift modulation, with 10 training samples in the low-data split and 100 in the higher-data split.The respective training, validation, and test proportions are 5%/10%/85% and 50%/10%/40%.Source: Experiments Setup · p. 6

Experimental setup

  • For fair comparison, PHIA and all baseline algorithms undergo hyperparameter searches on the validation set; grid search covers important hyperparameters for each model, and final selected sets are summarized in Table 2.The chosen hyperparameters yield the lowest loss on the validation set.Source: Experiments Setup · pp. 6–7, Table 2 · p. 6

Experimental comparison

  • PHIA’s modeled waveform is reported as closest to measurement among the top five benchmarks, indicating the strongest modeling precision in the plotted comparison.Figure 5 presents the modeled inductor-current waveforms.Source: Figure 5 · p. 7

Experimental Results

PHIA outperforms the reported benchmarks in both data splits, including in the low-data setting, and its modeled current waveform aligns closely with measurements. The experiments report MAE comparisons and evaluate performance across repeated runs.

Experimental Results

  • Each algorithm was run ten times with its optimal hyperparameters, and average MAE was reported for training, validation, and test sets.Reported P-values come from t-tests at a 0.02 significance level.Source: Experimental Results · p. 7
  • Across both data splits, PHIA outperforms conventional machine-learning, deep-learning, and state-of-the-art forecasting baselines.With 10 training samples, PHIA outperforms most benchmarks trained with 100 samples and aligns with CirNet-Only.Source: Experimental Results · p. 7
  • PHIA’s modeled iL waveform aligns more closely with measurements than the other models’ waveforms.Figure 5 visualizes the modeled waveform comparison.Source: Experimental Results · p. 7
  • 63.2% lower MAE than the second-best benchmark, TST, is reported for PHIA with a 5% training split (10 samples).The results summarize ten runs per algorithm and report average MAE for training, validation, and test sets.Source: Table 3 · p. 7, Experimental Results · p. 7

Interpretation

The interpretation section presents PHIA’s accuracy, workflow efficiency, scalability, and explainability as practical strengths. It connects these qualities to power-electronics applications, including renewable-energy systems and grid operations.

Efficiency Gains

  • The interpretation attributes efficiency gains to engineers reaching desired outcomes with fewer iterations and reduced computational resources.It also associates workflow optimization with lower operational costs than traditional methods.Source: Efficiency Gains · p. 7

Scalability

  • Performance improvements across metrics are presented as indicating scalability to more complex systems or broader application domains without proportional resource increases.The passage names computational and human resources as the resources that need not increase proportionally.Source: Scalability · p. 7

Explainability and Transparency

  • Hierarchical modeling is presented as improving explainability by aligning predictions with physical principles.The interpretation describes the resulting outputs as interpretable and trustworthy.Source: Explainability and Transparency · p. 7

Accuracy and Reliability

  • 23.7% lower MAE than TST is reported for PHIA with a 50% training split (100 samples), alongside the stated robustness of low-data performance.The paper links precision and reliability to operational performance, while noting that failures in real-world applications are a risk to reduce.Source: Accuracy and Reliability · p. 7, Table 4 · p. 8

Implications for Decision-Making

  • The paper connects reliable, precise performance metrics to decision-making in renewable-energy systems and power-grid operations.It identifies operational success and system stability as application concerns influenced by those metrics.Source: Implications for Decision-Making · p. 8

User Study

A user study with 20 industrial practitioners compared PHIA with analytical and manual design for a DAB converter TPS modulation task. All participants completed both tasks, with PHIA requiring substantially fewer measured time blocks.

User Study

  • The use case was designed to assess guidance in defining specifications, tailored design outcomes, and analysis of those outcomes.The study evaluated PHIA’s empirical performance in engineering efficiency for DAB converter modulation design.Source: User Study · p. 8
  • The study recruited 20 practitioners: 10 junior engineers with under three years’ experience and 10 senior engineers with over five years’ experience.Participants volunteered through research labs and collaborating external companies, and none had prior AI-assisted engineering-design experience.Source: User Study · p. 8
  • Participants used PHIA’s chat interface for one task variant, then standard analytical and manual design for the other, with working hours measured in 30-minute blocks.Both variants concerned designing a DAB converter’s TPS modulation strategy as a DC transformer for a DC-DC power grid.Source: User Study · p. 8
  • Junior engineers averaged 1.2 time blocks with PHIA versus 115.5 analytically, while seniors averaged 0.9 versus 30.5; reported efficiency gains were 96.3x and 33.9x, respectively.All 20 participants completed both assigned tasks successfully.Source: User Study · p. 8

Conclusion

The conclusion presents PHIA as an LLM-based agent for power-converter modulation design that combines physics-informed modeling with an LLM planning module. It reports accurate modeling in data-constrained settings and over 33-fold efficiency gains in the user study.

Conclusion

  • PHIA combines LLM reasoning with physical constraints to provide robust and interpretable power-converter modulation designs.The conclusion describes the system as an LLM-based, physics-informed autonomous agent.Source: Conclusion · p. 8
  • The hierarchical physics-informed neural-network architecture is presented as supporting accuracy and interpretability, while LLM planning provides an intuitive interface for engineers.The conclusion notes accurate power-converter modeling in data-constrained settings.Source: Conclusion · p. 8
  • The user study reports that engineers improved efficiency by over 33 times when using PHIA in practice.The conclusion identifies extending PHIA to other power-electronics applications and adding physics-based constraints as future work.Source: Conclusion · p. 8

Appendix A: Example Use Case

The use case shows PHIA collecting converter-design specifications through a chat and presenting the design outcome with text and figures. The example selects a dual active bridge strategy, prioritizes current stress, and provides modulation results for deployment.

Appendix A: Example Use Case

  • PHIA gathers a dual active bridge design request, then elicits the modulation strategy and prioritized objective through chat.The example selects TPS and current stress, respectively.Source: Appendix A · p. 11, Table 5 · p. 11
  • The conversation collects rated and operating conditions, asks the user to select an optimization algorithm, and confirms specifications before execution.The example provides PSO as the selected algorithm and lists the converter’s rated and operating values.Source: Appendix A · p. 11
  • The example presents current-stress modeling over inner phase shifts, compares optimal TPS with SPS across 100–1000 W, and identifies deployable modulation values.For 200 V input, 600 W power, and 160 V output, the best-performing phase-shift pair is marked on the modeling results.Source: Appendix A · p. 11
Loading 2411.14214v2…