arxiv: 2605.02133 · v1 · submitted 2026-05-04 · 💻 cs.LG

Recognition: unknown

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

Hongwei Jin , Keunju Song , Zeeshan Memon , Yijiang Li , Stefano Fenu , Hongseok Kim , Liang Zhao , Kibaek Kim

Authors on Pith no claims yet

Pith reviewed 2026-05-09 16:41 UTC · model grok-4.3

classification 💻 cs.LG

keywords AC optimal power flowsurrogate learningbenchmark suitepower gridmachine learningconstraint-aware trainingtopology generalizationfeasibility

0 comments

The pith

LUMINA-Bench supplies a standardized suite for testing whether AC optimal power flow surrogate models generalize to network topologies absent from training.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces LUMINA-Bench to tackle the fact that surrogate models for AC optimal power flow calculations frequently produce unreliable results on grids whose structure differs from the training data. It supplies data processing pipelines, training protocols, and evaluation metrics that cover both single-topology and multi-topology pretraining, transfer, and adaptation scenarios. The suite also compares plain regression losses against constraint-aware objectives such as augmented Lagrangian and violation-based penalties, tracking both prediction error and physical feasibility. By releasing the full framework, the work gives researchers a common yardstick for measuring progress toward surrogates that remain usable in operational what-if analyses on changing grids.

Core claim

We introduce LUMINA-Bench, a comprehensive benchmark suite for ACOPF surrogate learning covering multi-topology pretraining, transfer, and adaptation. The benchmark evaluates homogeneous and heterogeneous architectures under single- and multi-topology learning settings using unified metrics that capture both predictive accuracy and physics-informed constraint violations. We additionally compare constraint-aware training objectives, including MSE, augmented Lagrangian, and violation-based Lagrangian losses, to characterize accuracy-robustness trade-offs across settings.

What carries the argument

LUMINA-Bench, the open-sourced benchmark that standardizes data handling, model training, and evaluation for AC optimal power flow surrogates across multiple network topologies with metrics for both accuracy and constraint violations.

If this is right

Multi-topology pretraining followed by adaptation produces surrogates that maintain higher feasibility on new networks than single-topology training.
Violation-based Lagrangian losses reduce the rate of constraint violations at a modest cost in prediction accuracy compared with plain MSE.
Unified metrics that jointly track error and feasibility allow direct comparison of homogeneous and heterogeneous neural architectures across learning regimes.
Open release of the data pipelines and evaluation code removes a major barrier to reproducible progress on feasible OPF surrogates.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The benchmark could serve as a testbed for developing grid foundation models that treat topology variation as a core input rather than a distribution shift.
Results from the suite might guide selection of loss functions for surrogates deployed in day-ahead market clearing or contingency analysis where constraint satisfaction is non-negotiable.
Extending the framework to include time-varying loads or stochastic renewable injections would expose whether current architectures also generalize along the time dimension.

Load-bearing premise

That the selected accuracy metrics and constraint-aware losses will reliably flag models capable of producing feasible solutions on real grids whose topologies were never seen during training.

What would settle it

A surrogate that ranks highest on all LUMINA-Bench scores yet yields power-flow solutions that violate voltage or line limits when applied to an actual grid topology drawn from operating records outside the benchmark's dataset collection.

Figures

Figures reproduced from arXiv: 2605.02133 by Hongseok Kim, Hongwei Jin, Keunju Song, Kibaek Kim, Liang Zhao, Stefano Fenu, Yijiang Li, Zeeshan Memon.

**Figure 1.** Figure 1: Architecture comparison on single-topology (left) vs. multi-topology (right) training. Heteroge view at source ↗

**Figure 2.** Figure 2: Model error on held-out test data across cases 14, 30, 57,118, 500, 2000 with respect to total system view at source ↗

**Figure 3.** Figure 3: MAE contribution of individual buses with respect to their topological distance to reference view at source ↗

**Figure 4.** Figure 4: Loss function comparison across architectures under equal training budgets. Left panels: single view at source ↗

**Figure 5.** Figure 5: Cost–feasibility tradeoff with and without cost objective view at source ↗

**Figure 6.** Figure 6: Topology-normalized total constraint violation across increasing power-grid sizes. Results are view at source ↗

**Figure 7.** Figure 7: Constraint violation convergence on case500: fine-tuning vs. training from scratch view at source ↗

**Figure 8.** Figure 8: Training time vs. case size with and without mixed precision training(BF16) view at source ↗

**Figure 9.** Figure 9: Model error across categories with respect to node degree, across cases 14, 30, 57, 118, 500, 2000 view at source ↗

**Figure 10.** Figure 10: Hyperparameter optimization sensitivity analysis across 2M training samples. (a) Transformer view at source ↗

**Figure 11.** Figure 11: Relative training time for Transformer (left panel) and HGT (right panel) with AL and VBL loss view at source ↗

**Figure 12.** Figure 12: PCA of layer activations for samples from case118, labeled by node type, for GCN, Transformer, view at source ↗

**Figure 13.** Figure 13: PCA components of activation for the top layer of convolutions in HGT trained on AL (left) view at source ↗

**Figure 14.** Figure 14: Linear probing of HGT layers with respect to system load. The choice of AL loss induces view at source ↗

read the original abstract

AC optimal power flow (ACOPF) is foundational yet computationally expensive in power grid operations, driving learning-based surrogates for large-scale grid analysis. These surrogates, however, often fail to generalize across network topologies, a critical gap for deployment on grids not seen during training and for routine operational what-if studies. We introduce LUMINA-Bench, a comprehensive benchmark suite for ACOPF surrogate learning covering multi-topology pretraining, transfer, and adaptation. The benchmark evaluates homogeneous and heterogeneous architectures under single- and multi-topology learning settings using unified metrics that capture both predictive accuracy and physics-informed constraint violations. We additionally compare constraint-aware training objectives, including MSE, augmented Lagrangian, and violation-based Lagrangian losses, to characterize accuracy-robustness trade-offs across settings. Data processing, training, and evaluation frameworks are open-sourced as the LUMINA suite to support reproducibility and accelerate future research on feasibility-aware OPF surrogates.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

LUMINA-Bench gives the ACOPF surrogate field a multi-topology testbed with open code and unified accuracy-plus-violation metrics, which is the main thing worth noting.

read the letter

The paper's core contribution is a new benchmark suite that moves beyond the usual single-topology ACOPF surrogate experiments. It defines LUMINA-Bench to cover pretraining, transfer, and adaptation across multiple grids, then runs homogeneous and heterogeneous models under single- and multi-topology regimes. The authors also compare plain MSE against augmented Lagrangian and violation-based losses to map accuracy-robustness trade-offs, and they release the full data-processing, training, and evaluation stack. That open release is the practical part that matters most right now; anyone working on feasibility-aware OPF models can actually use the same splits and metrics instead of reinventing them. The experimental setup looks reproducible from the abstract description, and the focus on constraint violations alongside prediction error is a reasonable step for this domain. The work does not claim the chosen metrics will automatically pick models that work in live operations, which keeps the claims grounded. The main soft spot is that the specific topology selection and data splits are not yet visible in detail, so it is hard to judge how representative the benchmark is of real grid variability or how sensitive the reported trade-offs are to those choices. Minor gaps like that are common in first-version benchmarks and can be fixed in revision. This paper is aimed at the small but growing group of researchers who build or evaluate learning surrogates for power flow problems. Anyone who needs a shared testbed or wants to compare loss functions on the same footing will find it directly useful. It is the kind of infrastructure paper that deserves a serious referee rather than a desk reject, because the field benefits from standardized evaluation even if the particular metrics later get refined.

Referee Report

0 major / 3 minor

Summary. The manuscript introduces LUMINA-Bench, a benchmark suite for AC optimal power flow (ACOPF) surrogate learning. It covers multi-topology pretraining, transfer, and adaptation scenarios; evaluates homogeneous and heterogeneous neural architectures under single- and multi-topology settings; employs unified metrics capturing both predictive accuracy and physics-informed constraint violations; compares constraint-aware training objectives (MSE, augmented Lagrangian, and violation-based Lagrangian losses) to characterize accuracy-robustness trade-offs; and open-sources the associated data processing, training, and evaluation frameworks as the LUMINA suite.

Significance. If the benchmark construction and reported characterizations hold, the work is significant for the field of learning-based power system surrogates. Standardized multi-topology evaluation with explicit constraint-violation metrics addresses a recognized deployment barrier (generalization to unseen grids), while the open-sourced suite and loss-function comparisons can accelerate reproducible research on feasibility-aware models.

minor comments (3)

Abstract and title: the phrasing 'LUMINA: A Grid Foundation Model for Benchmarking' risks conflating a benchmark suite with a foundation model; explicit clarification of scope (benchmark definition and characterization rather than a new pretrained model) would improve precision.
Evaluation sections: while the abstract states that data splits, topology selection, and statistical significance are described, the manuscript would benefit from a dedicated subsection summarizing the exact criteria used for topology diversity (size, connectivity, load patterns) and the number of random seeds or statistical tests supporting the accuracy-robustness trade-off claims.
Metrics and losses: the unified metrics are described at a high level; adding a short table or pseudocode box that shows how constraint violations are aggregated across buses and time periods would aid reproducibility for readers implementing the benchmark.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive assessment of LUMINA-Bench, its significance for learning-based power system surrogates, and the recommendation for minor revision. No specific major comments were provided in the report.

Circularity Check

0 steps flagged

No significant circularity; benchmark definition is self-contained

full rationale

The manuscript introduces LUMINA-Bench as a benchmark suite for ACOPF surrogate learning, including multi-topology pretraining, transfer, adaptation, unified metrics (accuracy and constraint violations), and comparisons of loss functions (MSE, augmented Lagrangian, violation-based). No derivations, first-principles predictions, fitted parameters renamed as outputs, or load-bearing self-citations appear. The central claim is the construction and open-sourcing of the benchmark itself, which does not reduce to its own inputs by definition or equation. This is a standard benchmark paper with independent content in data processing, evaluation frameworks, and empirical characterization of trade-offs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, mathematical axioms, or invented entities are introduced; the contribution is an empirical benchmark definition resting on domain-standard assumptions about ACOPF physics and surrogate evaluation.

pith-pipeline@v0.9.0 · 5490 in / 1024 out tokens · 39763 ms · 2026-05-09T16:41:38.048208+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

40 extracted references · 8 canonical work pages · 3 internal anchors

[1]

Understanding intermediate layers using linear classifier probes

Guillaume Alain and Yoshua Bengio. Understanding intermediate layers using linear classifier probes. arXiv preprint arXiv:1610.01644, 2016

work page Pith review arXiv 2016
[2]

Strong np-hardness of ac power flows feasibility.Operations Research Letters, 47(6):494–501, 2019

Daniel Bienstock and Abhinav Verma. Strong np-hardness of ac power flows feasibility.Operations Research Letters, 47(6):494–501, 2019

2019
[3]

Grid structural characteristics as validation criteria for synthetic networks.IEEE Transactions on power systems, 32(4):3258–3265, 2016

Adam B Birchfield, Ti Xu, Kathleen M Gegner, Komal S Shetye, and Thomas J Overbye. Grid structural characteristics as validation criteria for synthetic networks.IEEE Transactions on power systems, 32(4):3258–3265, 2016

2016
[4]

Aug- mented lagrangian guided learning for the optimal power flow problem.IFAC-PapersOnLine, 58(13):50– 55, 2024

Sarra Bouchkati, Philipp Lutat, Luis B¨ ottcher, Florian Klein-Helmkamp, and Andreas Ulbig. Aug- mented lagrangian guided learning for the optimal power flow problem.IFAC-PapersOnLine, 58(13):50– 55, 2024

2024
[5]

Hammerla

Dan Busbridge, Dane Sherburn, Pietro Cavallo, and Nils Y. Hammerla. Relational graph attention networks, 2019

2019
[6]

Carpentier

J. Carpentier. Contribution ` a l’´ etude du dispatching ´ economique.Bulletin de la Soci´ et´ e Fran¸ caise des ´Electriciens, 3(8):431–447, 1962

1962
[7]

Physics-informed gradient estimation for accelerating deep learning-based ac-opf.IEEE Transactions on Industrial Informatics, 2025

Kejun Chen, Shourya Bose, and Yu Zhang. Physics-informed gradient estimation for accelerating deep learning-based ac-opf.IEEE Transactions on Industrial Informatics, 2025

2025
[8]

Initial estimate of AC optimal power flow with graph neural networks.Electric Power Systems Research, 234:110782, 2024

Amir Deihim, Despina Apostolopoulou, and Eduardo Alonso. Initial estimate of AC optimal power flow with graph neural networks.Electric Power Systems Research, 234:110782, 2024. 14

2024
[9]

Predicting AC optimal power flows: Combining deep learning and lagrangian dual methods

Ferdinando Fioretto, Terrence WK Mak, and Pascal Van Hentenryck. Predicting AC optimal power flows: Combining deep learning and lagrangian dual methods. InProceedings of the AAAI conference on artificial intelligence, volume 34, pages 630–637, 2020

2020
[10]

Optimal power flow: A bibliographic survey i: Formulations and deterministic methods.Energy systems, 3(3):221–258, 2012

Stephen Frank, Ingrida Steponavice, and Steffen Rebennack. Optimal power flow: A bibliographic survey i: Formulations and deterministic methods.Energy systems, 3(3):221–258, 2012

2012
[11]

Optimal power flow: A bibliographic survey ii: Non-deterministic and hybrid methods.Energy systems, 3(3):259–289, 2012

Stephen Frank, Ingrida Steponavice, and Steffen Rebennack. Optimal power flow: A bibliographic survey ii: Non-deterministic and hybrid methods.Energy systems, 3(3):259–289, 2012

2012
[12]

A physics-guided graph convolution neural network for optimal power flow.IEEE Transactions on Power Systems, 39(1):380–390, 2023

Maosheng Gao, Juan Yu, Zhifang Yang, and Junbo Zhao. A physics-guided graph convolution neural network for optimal power flow.IEEE Transactions on Power Systems, 39(1):380–390, 2023

2023
[13]

Foun- dation models for the electric power grid.Joule, 8(12):3245–3258, 2024

Hendrik F Hamann, Blazhe Gjorgiev, Thomas Brunschwiler, Leonardo SA Martins, Alban Puech, Anna Varbella, Jonas Weiss, Juan Bernabe-Moreno, Alexandre Blondin Mass´ e, Seong Lok Choi, et al. Foun- dation models for the electric power grid.Joule, 8(12):3245–3258, 2024

2024
[14]

Alternative learning architecture for solving ac-opf via supervised relaxation and cross encoder

Doan Thanh Hien, Keunju Song, Kibaek Kim, and Hongseok Kim. Alternative learning architecture for solving ac-opf via supervised relaxation and cross encoder. InNeurIPS Workshop on GPU-Accelerated and Scalable Optimization, 2025

2025
[15]

Heterogeneous graph transformer

Ziniu Hu, Yuxiao Dong, Kuansan Wang, and Yizhou Sun. Heterogeneous graph transformer. InPro- ceedings of the web conference 2020, pages 2704–2710, 2020

2020
[16]

Wanjun Huang, Xiang Pan, Minghua Chen, and Steven H. Low. DeepOPF-V: Solving AC-OPF problems efficiently.IEEE Transactions on Power Systems, 37(1):800–803, 2022

2022
[17]

Accelerated computation and tracking of ac optimal power flow solu- tions using gpus

Youngdae Kim and Kibaek Kim. Accelerated computation and tracking of ac optimal power flow solu- tions using gpus. InWorkshop Proceedings of the 51st International Conference on Parallel Processing, pages 1–8, 2022

2022
[18]

Semi-Supervised Classification with Graph Convolutional Networks

TN Kipf. Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907, 2016

work page internal anchor Pith review arXiv 2016
[19]

Gpu-accelerated sequential quadratic programming algorithm for solving acopf

Bowen Li and Kibaek Kim. Gpu-accelerated sequential quadratic programming algorithm for solving acopf. In2024 IEEE 63rd Conference on Decision and Control (CDC), 2024

2024
[20]

Topology-aware graph neural networks for learning feasible and adaptive ac-opf solutions.IEEE Transactions on Power Systems, 38(6):5660–5670, 2022

Shaohui Liu, Chengyang Wu, and Hao Zhu. Topology-aware graph neural networks for learning feasible and adaptive ac-opf solutions.IEEE Transactions on Power Systems, 38(6):5660–5670, 2022

2022
[21]

Opfdata: Large-scale datasets for ac optimal power flow with topological perturbations.arXiv preprint arXiv:2406.07234, 2024

Sean Lovett, Miha Zgubic, Sofia Liguori, Sephora Madjiheurem, Hamish Tomlinson, Sophie Elster, Chris Apps, Sims Witherspoon, and Luis Piloto. Opfdata: Large-scale datasets for ac optimal power flow with topological perturbations.arXiv preprint arXiv:2406.07234, 2024

work page arXiv 2024
[22]

Multi-agent trajectory prediction with het- erogeneous edge-enhanced graph attention network.IEEE Transactions on Intelligent Transportation Systems, 23(7):9554–9567, 2022

Xiaoyu Mo, Zhuyu Huang, Yang Xing, and Chen Lv. Multi-agent trajectory prediction with het- erogeneous edge-enhanced graph attention network.IEEE Transactions on Intelligent Transportation Systems, 23(7):9554–9567, 2022

2022
[23]

Momoh, R

J.A. Momoh, R. Adapa, and M.E. El-Hawary. A review of selected optimal power flow literature to
[24]

nonlinear and quadratic programming approaches.IEEE Transactions on Power Systems, 14(1):96–104, 1999

i. nonlinear and quadratic programming approaches.IEEE Transactions on Power Systems, 14(1):96–104, 1999

1999
[25]

Madncl: a gpu implementation of algorithm ncl for large-scale, degenerate nonlinear programs.arXiv preprint arXiv:2510.05885, 2025

Alexis Montoison, Fran¸ cois Pacaud, Michael Saunders, Sungho Shin, and Dominique Orban. Madncl: a gpu implementation of algorithm ncl for large-scale, degenerate nonlinear programs.arXiv preprint arXiv:2510.05885, 2025

work page arXiv 2025
[26]

Physics-informed neural networks for ac optimal power flow.Electric Power Systems Research, 212:108412, 2022

Rahul Nellikkath and Spyros Chatzivasileiadis. Physics-informed neural networks for ac optimal power flow.Electric Power Systems Research, 212:108412, 2022. 15

2022
[27]

Optimal power flow using graph neural net- works

Damian Owerko, Fernando Gama, and Alejandro Ribeiro. Optimal power flow using graph neural net- works. InProc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5930–5934, Barcelona, Spain, 2020

2020
[28]

An aug- mented lagrangian method on gpu for security-constrained ac optimal power flow.arXiv preprint arXiv:2510.13333, 2025

Fran¸ cois Pacaud, Armin Nurkanovi´ c, Anton Pozharskiy, Alexis Montoison, and Sungho Shin. An aug- mented lagrangian method on gpu for security-constrained ac optimal power flow.arXiv preprint arXiv:2510.13333, 2025

work page arXiv 2025
[29]

Xiang Pan, Minghua Chen, Tianyu Zhao, and Steven H. Low. DeepOPF: A feasibility-optimized deep neural network approach for AC optimal power flow problems.IEEE Systems Journal, 17(1):673–683, 2023

2023
[30]

Xiang Pan, Wanjun Huang, Minghua Chen, and Steven H. Low. DeepOPF-AL: Augmented learning for solving AC-OPF problems with a multi-valued load-solution mapping. InThe 14th ACM International Conference on Future Energy Systems (e-Energy ’23), pages 391–396, Orlando, FL, USA, 2023

2023
[31]

CANOS : A fast and scalable neural AC - OPF solver robust to N-1 perturbations

Luis Piloto, Sofia Liguori, Sephora Madjiheurem, Miha Zgubic, Sean Lovett, Hamish Tomlinson, Sophie Elster, Chris Apps, and Sims Witherspoon. Canos: A fast and scalable neural ac-opf solver robust to n-1 perturbations.arXiv preprint arXiv:2403.17660, 2024

work page arXiv 2024
[32]

Accelerating optimal power flow with gpus: Simd abstraction of nonlinear programs and condensed-space interior-point methods.Electric Power Systems Research, 236:110651, 2024

Sungho Shin, Mihai Anitescu, and Fran¸ cois Pacaud. Accelerating optimal power flow with gpus: Simd abstraction of nonlinear programs and condensed-space interior-point methods.Electric Power Systems Research, 236:110651, 2024

2024
[33]

Scalable multi- period ac optimal power flow utilizing gpus with high memory capacities

Sungho Shin, Vishwas Rao, Michel Schanen, D Adrian Maldonado, and Mihai Anitescu. Scalable multi- period ac optimal power flow utilizing gpus with high memory capacities. In2024 Open Source Modelling and Simulation of Energy Systems (OSMSES), pages 1–6. IEEE, 2024

2024
[34]

Graph Attention Networks

Petar Veliˇ ckovi´ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks.arXiv preprint arXiv:1710.10903, 2017

work page internal anchor Pith review arXiv 2017
[35]

Data-driven ac optimal power flow with physics-informed learning and calibrations, 2024

Junfei Wang and Pirathayini Srikantha. Data-driven ac optimal power flow with physics-informed learning and calibrations, 2024

2024
[36]

How Powerful are Graph Neural Networks?

Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826, 2018

work page internal anchor Pith review arXiv 2018
[37]

Topology-transferable physics-guided graph neural network for real-time optimal power flow

Mei Yang, Gao Qiu, Junyong Liu, Youbo Liu, Tingjian Liu, Zhiyuan Tang, Lijie Ding, Yue Shui, and Kai Liu. Topology-transferable physics-guided graph neural network for real-time optimal power flow. IEEE Transactions on Industrial Informatics, 20(9):10857–10872, 2024

2024
[38]

Graph transformer networks.Advances in neural information processing systems, 32, 2019

Seongjun Yun, Minbyul Jeong, Raehyun Kim, Jaewoo Kang, and Hyunwoo J Kim. Graph transformer networks.Advances in neural information processing systems, 32, 2019

2019
[39]

Qcqp-net: Reliably learning feasible alternating current optimal power flow solutions under constraints

Sihan Zeng, Youngdae Kim, Yuxuan Ren, and Kibaek Kim. Qcqp-net: Reliably learning feasible alternating current optimal power flow solutions under constraints. In6th Annual Learning for Dynamics & Control Conference, volume 242, 2024

2024
[40]

Chuxu Zhang, Dongjin Song, Chao Huang, Ananthram Swami, and Nitesh V. Chawla. Heterogeneous graph neural network. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD), pages 793–803, 2019. The submitted manuscript has been created by UChicago Argonne, LLC, Operator of Argonne Na- tional Laboratory (“Argo...

2019