Session Chair: Hendrik F. Hamann
Brookhaven National Laboratory and Stony Brook University

This session provides an update on the state of the art of foundation models for power systems. Speakers will review current GridFM results, separating proven capabilities from promising ideas. Topics include architectures, training, fine-tuning, evaluation, generalizability, adaptability, data efficiency, accuracy, topology robustness, shared benchmarks, roadmaps, and collaboration.

GENCO — A Unified Neural Solver Embedded in a Development Framework for Steady-State Grid Analysis

Speaker: Thomas Brunschwiler – IBM Research

We present GENCO (GEometric Neural Corrective Optimizer), a unified neural solver for steady-state transmission grid analysis that addresses power flow (PF), optimal power flow (OPF), and state estimation (SE) within a single architecture and shared grid representation. To accelerate advances in neural power-system solvers, we introduce the open-source GridFM Development Framework, which standardizes synthetic data generation and model training in a low-code environment, together with large-scale datasets containing millions of PF and OPF scenarios across diverse grid topologies.

We evaluate GENCO on state-of-the-art PF and OPF benchmarks against leading neural and classical solvers, and validate SE on real-world Hydro-Québec SCADA data. For large-scale PF, GENCO recovers the full AC operating state—including voltage magnitudes and reactive power unavailable from DC-PF—while achieving DC-PF-level active power-balance residuals and up to 30× speedups over Newton–Raphson, at only 2× the runtime of DC-PF. For OPF, GENCO achieves up to 85× speedups over IPOPT while improving feasibility, optimality, and runtime over DC-OPF. For SE, GENCO is more robust than classical weighted least squares to noisy measurements and grid parameters and returns high-quality estimates even when WLS fails to converge.

Together, GENCO and the GridFM Development Framework enable scalable steady-state grid analysis, reduce integration effort, and lower barriers to developing neural grid solvers, representing a key step toward Grid Foundation Models.

LUMINA: Open Weights, Many Tasks, and the Road to Scale

Speaker: Kim Kibaek – Argonne National Laboratory

Foundation models for power systems are easy to announce and hard to hand over. This talk reports on LUMINA — Large-scale Unified Model for Intelligent Grid Applications — the U.S. DOE multi-lab effort to build grid foundation models that others can download, run, and retrain. The first, LUMINA-2M, is a 2.4-million-parameter heterogeneous graph transformer trained as an alternating-current optimal power flow (AC-OPF) surrogate, published on Hugging Face under Apache-2.0 alongside two open repositories: a PyTorch training framework and a lightweight inference package. Training uses a physics-informed augmented Lagrangian loss over full-topology and N-1 contingency data; checkpoints, configurations, and the constraint-violation evaluator are all public. Two directions follow. The first is scale: I will present a training-scalability study on NERSC’s Perlmutter and what it implies for the larger models now coming through the same release path. The second is breadth: extending one shared backbone from AC-OPF to security-constrained unit commitment, transient stability, reliability assessment, and distribution system state estimation as task heads on a single model, rather than a separate model per application. I will close on why this is worth doing. Open weights, open training code, and reproducible evaluation let results accumulate rather than restart, turning a model into shared infrastructure. No single group will build the grid foundation model. Community efforts like GridFM are how the field gets there, and Argonne intends to keep releasing into that commons.

Introducing gridfm-app-distribution: GridFM for power distribution applications

Mikka Kisuule – INESC-TEC

The talk will focus on a presentation of the gridfm-app-distribution repo, which GridFM recently launched publicly. The objective of the talk is to introduce the repo, discuss the current progress on the voltage loss detection (VLD) task and invite the audience to use it/contribute. The talk will be divided into 3 parts:
The first part will provide a context for the repo, presenting the challenges it seeks to address and the motivation for a new model tailored for distribution applications. This part will also discuss the topology awareness of GridFMs, how the VLD task reconstructs bus voltage collapse following a network contingency, and the model architecture.
In the second part, we will provide an update on the current progress with the loss function, how the VLD task works and what the model learns during training. The results generated so far and the challenges overcome will also be discussed.
Finally, we will discuss the way forward for the repo, highlighting the limitations of the model, such as the need for diverse training datasets, computing capacity for training large datasets, and model tuning. The prospective distribution applications to be explored, such as nodal reception and hosting capacity and reconstruction of impedance and short-circuit levels, will also be discussed.