GridAI

AI Foundation Models & Agents for the Power Grid

A research initiative advancing machine learning for power system optimization, analysis, and operations — from graph-based foundation models to LLM-orchestrated agentic workflows.

Argonne National Laboratory LDRD Funding ModCon Genesis Mission
Scroll ↓
The Project

Intelligent AI for a Resilient Power Grid

GridAI brings modern machine learning to the hardest problems in power system operations. The project is organized around two complementary product lines: LUMINA, a family of graph-based foundation models, and GridMind, an agentic framework that reasons about grid risk and operations.

◈
Foundation Model

LUMINA

Large-scale Unified Model for Intelligent Grid Applications. A graph neural network foundation model for power grid applications — multi-task, multi-dataset, and transferable across grid topologies. AC Optimal Power Flow (ACOPF) is the current primary focus.

Learn more →
◆
Agentic Framework

GridMind

An LLM-orchestrated, LangGraph-based agentic framework for intelligent grid applications. Multi-agent workflows tackle weather-driven risk assessment, power system analysis, and operational decision support through an interactive interface.

Learn more →
LUMINA

A Grid Foundation Model that Transfers

LUMINA is a multi-task, multi-dataset foundation model for power grid applications. Its defining strength is generalization: a single graph-based model that transfers across grid topologies, scales, and tasks — moving beyond one-model-per-grid. AC Optimal Power Flow (ACOPF) is our current primary focus, with the framework designed to extend to a broader family of grid optimization and analysis tasks.

  • Topology-transferable — generalizes to unseen power networks without retraining
  • Multi-task & multi-dataset — one backbone across tasks and grid scales (14 → 13,659 bus)
  • Heterogeneous graph transformer over power-network structure
  • Physics-informed objectives for constraint feasibility
  • ACOPF today, extensible to the broader grid optimization & analysis family
  • LUMINA-Bench — open benchmark for ACOPF surrogate learning
LUMINA architecture
LUMINA architecture: parallel edge and node graph transformers (Q/K/V attention, Add & Norm, Scale, Softmax, Matmul) producing edge and node embeddings that combine into the prediction.

Parallel edge- and node-oriented transformers produce heterogeneous embeddings (Bus / Gen / Load / Shunt, AC-Line / Transformer) that are fused for prediction.

GridMind demo — agentic workflow in action

GridMind

Agentic Workflows for Grid Operations

GridMind is an LLM-powered, LangGraph-based agentic framework. An orchestrator routes tasks to specialized agents that perform power system analysis, contingency and cascading-failure reasoning, and weather-driven risk assessment — calling neural solvers like LUMINA when needed.

  • LLM orchestrator routing to deterministic, specialized executors
  • Weather-driven power grid risk assessment
  • Power system analysis & operations support
  • Interactive interface with reproducible, content-addressed artifacts
Research

Publications

Peer-reviewed papers and preprints from the GridAI team. Filter by component.

Aug 2026
KDD 2026 · AI for Science

Towards Systematic Generalization for Power Grid Optimization Problems

A unified learning framework that jointly solves AC Optimal Power Flow and Security-Constrained Unit Commitment through a shared graph backbone, with physics-informed training and a consensus mechanism for dispatch — achieving strong cross-topology transfer without retraining.

Aug 2026
KDD 2026 · Workshop

GridMind: An Agentic Workflow for Weather-Driven Power Grid Risk Assessment

An agentic workflow that couples weather intelligence with power system reasoning to assess grid risk, orchestrating specialized agents for end-to-end risk analysis.

May 2026
ICLR 2026 · Workshop

LUMINA: Foundation Models for Topology Transferable ACOPF

Introduces LUMINA foundation models that learn ACOPF surrogates transferable across grid topologies, enabling a single model to generalize to unseen power networks.

May 2026
arXiv preprint

LUMINA: A Grid Foundation Model for Benchmarking AC Optimal Power Flow Surrogate Learning

A comprehensive benchmark (to be renamed LUMINA-Bench) evaluating homogeneous and heterogeneous neural architectures for ACOPF under single- and multi-topology settings, with unified metrics for accuracy and physics-constraint satisfaction.

Sep 2025
SC25 · Workshop

GridMind: LLMs-Powered Agents for Power System Analysis and Operations

The earlier release of GridMind — LLM-powered agents for power system analysis and operations, establishing the shared agentic framework that continues to evolve.

Open Source

Packages & Resources

Models, SDKs, and tools released by the GridAI team.

Collaboration

Team & Acknowledgments

GridAI is led by Argonne National Laboratory and carried out with a broad network of national laboratories, universities, and industry partners.

Laboratories & Universities

Argonne National Laboratory
Los Alamos National Laboratory
National Laboratory of the Rockies
Oak Ridge National Laboratory
Pacific Northwest National Laboratory
EPRI
Emory University
Massachusetts Institute of Technology
Harvard University

Industry Partners

Google
Amazon Web Services
NVIDIA
AMD
Siemens Energy
GE Vernova
Aveva
Tennessee Valley Authority
Bonneville Power Administration
American Electric Power
Southern Company
Exelon
Electric Reliability Council of Texas (ERCOT)
ISO New England
NRECA

Get in touch

For questions and collaboration inquiries, contact Kibaek Kim, Argonne National Laboratory — kimk@anl.gov