A closed-loop research workflow for power electronics, combining graph-based modeling, AI agents with direct tool access, and switched-circuit simulation.
ELITE Grid Research Lab
This repository describes a working methodology in which AI agents carry out the complete research cycle for power-electronic converters: selecting the topic, designing and running experiments, independently verifying the results, evaluating them against engineering standards, and archiving the outcome. The human researcher supervises the process, sets the standards, and makes the final decisions.
The workflow at a glance: a closed agent loop under human supervision, running on a graph-theoretic backbone — the knowledge graph grounds what to study, circuit graphs define what gets built and simulated, and the provenance graph records everything that happens.
The workflow combines three components. Each is established on its own; the practical capability comes from using them together.
Language models are unreliable when they work directly on unstructured material such as prose, slides, and informal notes. Graphs provide a uniform formal representation that can be enumerated, counted, and tested. This workflow uses graphs in three places:
- Circuit graphs. Converters are modeled as graphs (components as edges, junctions as nodes), so topology derivation and design translation become well-defined operations whose results can be checked mechanically.
- A knowledge graph. The laboratory's accumulated knowledge — publications, experiment reports, and verified textbook facts — is indexed as a graph, so retrieval is traceable to its sources.
- A provenance graph. Decisions, experiments, and rules are recorded as a graph, so every conclusion can be traced back to the record that produced it.
The graph-theoretic foundation is developed in the laboratory's monograph, Graph Theory in Power Converters (Wiley, forthcoming; see References).
Through the Model Context Protocol (MCP), the agents operate professional engineering software directly: they build and run simulations, process measurement data, and execute large enumerations in parallel. MCP is not tied to any one tool: the workflow is tool-agnostic, and the MATLAB/Simulink connection used here — via MathWorks' Simulink Agentic Toolkit — is simply the one we happen to run. MCP support across professional engineering software is developing quickly (circuit and real-time/HIL simulators, measurement and EDA tools, and more); the growing landscape is surveyed in Awesome MCP for Power Engineering. New tool connectors and collaboration are welcome. The work is divided among separate roles. A research agent performs the study; a first review agent independently reproduces and recomputes every reported number; a second review agent examines the interpretation for errors. Findings from review are recorded and converted into permanent automated checks.
Left: a language model used as an advisor — every iteration stalls at a human relay (building the model, reading waveforms, retyping numbers). Right: this workflow — the agent operates MATLAB/Simulink directly through MCP, parses the signals itself, and checks the results against pass/fail criteria fixed before the run.
Every claim is checked against switched-circuit simulation with device models: waveforms, harmonic content, and power balance. Acceptance criteria are fixed before an experiment runs, and simulation results — not model assertions — decide whether a claim is accepted.
In combination: graphs make the subject matter and the knowledge base machine-checkable, MCP gives the agents direct access to engineering tools, and simulation provides the physical reference. The result is a research cycle that completes in hours, with every reported number traceable to an archived result file.
Simulated three-phase waveforms of a converter design and its machine-translated twin: phase-voltage PWM of the voltage-source converter (three-level) and phase-current PWM of the resulting current-source converter (five-level). Both models were built and simulated in MATLAB/Simulink by the agents.
 |
 |
Agent-built Simulink/Simscape models: the voltage-source converter (left) and its current-source counterpart (right), constructed block-by-block from the graph representation — not drawn by hand.
 |
 |
Left: the graph representation behind both circuits — the same graph read two ways. Right: closed-loop start-up of the five-level current-source inverter from zero stored energy (see the case study for details).
The laboratory knowledge graph that grounds agent retrieval — publications, experiment reports, and verified textbook facts indexed as one connected structure (shown unlabeled).
A slice of the provenance graph: the documented causal chain behind one research finding — problem claims, experiment arcs, supervisor rulings, incident reports, and the automated rules they produced. Every edge is re-derivable from the archived records.
docs/presentation.html |
A 20-slide overview (self-contained HTML; download and open, navigate with arrow keys). Written for engineers without prior background in AI or graph theory. |
case-studies/five-level-csc-mpc/ |
Case study: a five-level current-source inverter taken from a paper topology to grid-quality closed-loop operation in one working day, including the controller revisions that failed. |
methodology/ |
The research pipeline, the supporting tools, and the governance process. |
docs/agentic_ai_matlab_simulink_mcp_tutorial_en.md |
Beginner-friendly tutorial: set up an AI coding agent to operate MATLAB/Simulink directly through MCP — the same connection used in this workflow. No prior MCP knowledge assumed. |
- Yuzhuo Li and Yunwei Li, Graph Theory in Power Converters: From Fundamentals to Applications, Wiley, forthcoming (ISBN 978-1-394-22229-2). The graph-theoretic foundation of this workflow.
- Y. Li, J. Kuprat, Y. Li, and M. Liserre, "Graph-Theory-Based Derivation, Modeling, and Control of Power Converter Systems," IEEE Journal of Emerging and Selected Topics in Power Electronics, 2022.
- R. W. Erickson and D. Maksimović, Fundamentals of Power Electronics, 3rd ed., Springer, 2020. The classical reference used for machine-verified textbook facts.
- IEEE Std 1547-2018, Standard for Interconnection and Interoperability of Distributed Energy Resources. Acceptance limit for grid-current distortion.
- IEEE Std 519, Recommended Practice and Requirements for Harmonic Control in Electric Power Systems.
- J. Jumper et al., "Highly accurate protein structure prediction with AlphaFold," Nature 596, 2021.
- D. A. Boiko et al., "Autonomous chemical research with large language models," Nature 624, 2023.
- A. Merchant et al., "Scaling deep learning for materials discovery," Nature 624, 2023.
- Model Context Protocol (MCP), the open interface through which the agents operate MATLAB/Simulink and other software. The MATLAB/Simulink side is provided by the Simulink Agentic Toolkit and the MATLAB MCP Core Server; a broader survey of MCP servers for engineering tools is collected in Awesome MCP for Power Engineering.
This is an ongoing project. The methodology and its supporting source code — the graph tooling, the agent harness, and the governance pipeline — are under active development and not yet public. Components will be shared selectively as they mature; this repository will be updated as that happens.
Content (documents, figures, slides) © ELITE Grid Research Lab, licensed under CC BY-NC-ND 4.0. Cite as: ELITE Grid Research Lab, "Graph-Driven Agentic Research Automation," 2026.