Scaling Laws for Physics-Aware ACOPF Surrogate Learning
Yijiang Li, Emon Dey, Stefano Fenu, Massimiliano Lupo Pasini, Teja Kuruganti, Kibaek Kim
cs.LG
Quick verdict
Scaling-law study of ACOPF surrogates: AL halves violation growth vs MSE at ~10× training cost; valuable but single-seed, cross-machine, one architecture.
Summary
This paper presents the first systematic empirical scaling-law characterization of physics-aware training for AC optimal power flow (ACOPF) surrogates. Using the LUMINA suite and public OPFData with a heterogeneous graph transformer (HGT) backbone, the authors sweep six model sizes (100M–671M parameters) and five dataset sizes (48K–240K samples) on case2000 under both MSE regression and an augmented Lagrangian (AL) objective, and separately compare both objectives on eight topologies from case30 to case10000 at a fixed budget. Fitted power laws show MSE prediction loss dominated by model capacity (joint exponents 0.949/0.596) while the AL composite score is balanced across both axes (0.284/0.292). Across topologies, constraint violation grows with network size at exponent 1.196±0.064 under MSE versus 0.513±0.068 under AL, with the largest gap on active power balance. On matched Frontier hardware at approximately equal prediction loss, AL reduces normalized violation 29.5× for 9.55× more wall-clock time with only ~30 MiB extra memory. Main limitations: the model–data grid uses a single random seed and trains the two objectives on different supercomputers (Aurora vs Frontier); the equal weighting of the AL composite metric is unjustified; hyperparameters were not re-tuned per scale; and only one architecture on synthetic data was tested.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Both MSE and AL training losses follow power laws in model and dataset size on case2000, with joint exponents (0.949, 0.596) for MSE and (0.284, 0.292) for the AL composite score.
- MSE scaling is dominated by model capacity while AL improves in a balanced way with model and data, so the training objective changes the rate of improvement with scale, not just the operating point.
- Constraint violation grows with network size at exponent 1.196±0.064 under MSE versus 0.513±0.068 under AL, i.e., physics-aware training roughly halves the feasibility degradation rate across case30–case10000.
- At matched prediction loss on identical Frontier hardware (671M model, 192K shard), AL reduces normalized violation 29.5× for 9.55× the elapsed training time with only ~30 MiB extra peak memory.
- AL's overhead falls entirely on training time, leaving the memory ceiling — and thus the maximum trainable model size on a given system — unchanged.
Strengths
Fills a genuine gap with a novel scaling axis
Sec. II-A documents that scientific scaling studies use only supervised objectives and that physical system size is an unstudied axis; the paper delivers both, fitting model–data laws (Eqs. 10–16) and a violation-vs-network-size trend across case30–case10000 (Sec. V-B).
Leadership-class execution with precise cost accounting
100M–671M parameter HGT runs on 256 Aurora nodes and 384 Frontier nodes (Sec. IV-D); Table I reports throughput, elapsed time, per-constraint-type violations, and peak memory for equal-sample and matched-loss settings separately.
Same-platform cost comparison avoids its own confound
Honest scoping and per-constraint decomposition
Weaknesses
Single seed for the entire model–data grid
Cross-machine confound in the main scaling comparison
Unjustified composite metric for AL
Frozen hyperparameters conflate objective with tuning depth
Fixed small budget for the cross-topology trend
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks whether the choice of training objective — plain regression (MSE) versus physics-aware training with an augmented Lagrangian (AL) — changes *how* ACOPF surrogate quality scales, not just where it lands. The authors position this against two documented gaps (Sec. II-A, II-C): existing scientific scaling studies use only supervised objectives, and no prior work characterizes constraint satisfaction as a function of physical network size. Three deliverables follow: (1) model–data scaling laws on case2000 (Sec. V-A), (2) violation-versus-network-size trends across eight topologies from case30 to case10000 (Sec. V-B), and (3) a matched-hardware cost/benefit analysis (Sec. V-C, Table I). The central claims are quantitatively stated: MSE loss is capacity-dominated (joint exponents α_N=0.949, β_D=0.596, Eq. 13) while the AL composite score is balanced (0.284, 0.292, Eq. 16); violation grows with network size at k=1.196±0.064 under MSE versus 0.513±0.068 under AL; and at matched prediction loss AL delivers a 29.5× violation reduction for 9.55× the wall-clock time with ~30 MiB extra memory. The conclusion honestly flags that exponents are local trends over the tested range.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the main finding of this paper?
For neural AC optimal power flow (ACOPF) surrogates, the training objective changes how quality scales. Under plain MSE the loss is mostly limited by model size (joint exponents 0.949 for model vs 0.596 for data), while under the augmented Lagrangian (AL) physics-aware objective model and data contribute about equally (0.284/0.292). Constraint violation grows with network size roughly twice as fast under MSE (exponent ~1.20) as under AL (~0.51), and at matched prediction loss on identical hardware AL cut violations ~30× at ~10× the training time.
Does physics-aware (augmented Lagrangian) training cost more than plain MSE?
Yes, but only in time. On matched Frontier hardware with a 671M-parameter HGT model, AL ran about 2.4× slower per sample and needed about 9.55× more wall-clock time to reach MSE's prediction-loss level, yet peak memory increased by only ~30 MiB (~0.1%), so it does not reduce the largest model that fits on a given system.
What data and models were used in the study?
The public OPFData dataset of solved ACOPF instances, a Heterogeneous Graph Transformer (HGT) backbone from the LUMINA benchmark suite, model sizes from 100M to 671M parameters, and dataset sizes from 48K to 240K training samples on the case2000 grid, plus eight topologies from case30 to case10000 for the network-size analysis.
What are the main limitations of this scaling-law study?
The model–data sweep uses a single random seed and trains the two objectives on different supercomputers (MSE on Aurora, AL on Frontier); the AL composite metric sums prediction loss and violation with an unjustified equal weighting; hyperparameters were not re-tuned per scale; and only one architecture (HGT) on synthetic OPFData was tested, so the fitted exponents are local trends valid over the tested range.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–4 are referenced but their plotted points could not be inspected from the text extraction; all figure-level numbers cited here come from in-text equations and Table I.
- No code or explicit data-release statement appears in the paper; OPFData and LUMINA are cited as public resources but their release status was not verifiable here.
- Several key citations are 2026 arXiv preprints (e.g., LUMINA, arXiv:2605.02133); their content could not be checked against this paper's claims.
- Node-level utilization, effective local batch decomposition, and dataloader behavior across Aurora vs Frontier are not fully specified, limiting assessment of the cross-machine comparability argument.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.