Pith. sign in

REVIEW 4 major objections 4 minor 32 references

Single-task GNNs overfit grid topology; joint multi-task training achieves zero boundary violations on unseen grids.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 03:45 UTC pith:JJCXMNH4

load-bearing objection Interesting topology-overfitting pattern, but the 0% BVR headline is confounded by the boundary loss only MxGPS receives. the 4 major comments →

arxiv 2607.13763 v2 pith:JJCXMNH4 submitted 2026-07-15 cs.LG cs.AI

MxGPS: Multiplex Graph Transformers for a Power Grid Foundation Model

Charilaos Papaioannou , Ioannis Tsantilas , Dimitris Giannakakos , Vasilis Michalakopoulos , Sotiris Pelekis , Vangelis Marinakis , Arsam Aryandoust , Antonello Monti
show 4 more authors
Ricardo J. Bessa Perdo P. Vergara Jochen Cremer Elissaios Sarmas
This is my paper
classification cs.LG cs.AI
keywords topology overfittinggraph neural networkgraph transformermulti-task learningpower flowstate estimationzero-shot generalizationpower grid foundation model
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper identifies a systematic failure mode it calls topology overfitting: graph neural networks fine-tuned on a single power-grid task achieve low error on the training grids but encode relational structure particular to those grids, and degrade by 190% to 1400% in error when tested on unseen topologies. To counter this, it introduces MxGPS, a multiplex graph transformer with a shared node encoder and two task-specialised branches, jointly trained on static state estimation and AC power flow. In a sliding-window cross-validation over eight IEEE test systems, MxGPS reaches a 0% boundary violation rate on all four zero-shot power-flow topologies and degrades by only 39% in power-flow error, while keeping state-estimation calibration intact. The paper argues that the joint gradient signal prevents the shared encoder from overfitting to any single task's topology-specific structure, making multi-task training a structural mechanism for topology-agnostic generalisation rather than a scaling fix.

Core claim

The central discovery is that topology overfitting is a structural failure of single-task GNN fine-tuning: the models with the lowest in-distribution power-flow error degrade the most under topology shift (GNS by 790%, GPS by over 1300%), while MxGPS degrades by only 39% and posts 0% boundary violations on all four zero-shot test grids. The paper attributes this to the joint SSE+PF training objective, which forces a shared encoder to satisfy complementary gradient signals simultaneously, preventing it from fitting topology-specific relational patterns. This is offered as direct evidence that generalisation to unseen grids is governed by the training objective rather than model capacity.

What carries the argument

MxGPS is a multiplex graph transformer: each of K task-specialised GPS branches (here K=2, for SSE and PF) runs over the same physical graph and shares a node encoder. A learnable task token conditions each branch, and an optional cross-branch attention module exchanges per-node representations between branches. The training loss combines per-task supervised errors with a physics-informed Bus Injection Model residual and a boundary penalty on voltage magnitudes. The carrying mechanism is the shared encoder's gradient signal: because it must satisfy both SSE and PF losses, it cannot encode relational structure that serves only one task, acting as an implicit regulariser against topology overf

Load-bearing premise

The claim that multi-task joint training is the cause of the zero-shot robustness rests on MxGPS being the only model trained with the physics and boundary losses; if those losses alone produce the same effect in a single-task model, the mechanism collapses.

What would settle it

Train a single-task GPS with the same physics loss and boundary penalty (λ_phys=0.01, λ_bnd=0.01) under the same protocol; if it also achieves 0% BVR and ~39% degradation on the zero-shot topologies, the multi-task mechanism is not the cause. Conversely, removing the boundary penalty from MxGPS while keeping joint training and observing BVR rise would support the penalty as the driver.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If topology overfitting is as described, then low in-distribution error is an unreliable predictor of zero-shot performance for single-task GNNs on power grids; foundation-model evaluation should include unseen-topology degradation and boundary violations.
  • Joint multi-task training on SSE and PF yields 0% boundary violations on all zero-shot topologies tested, implying operational safety constraints can be satisfied without explicit per-topology training.
  • The same GPS backbone trained single-task degrades over 14x under topology shift, while the multi-task variant degrades 39%, so the improvement is attributed to the training objective, not architecture or scale.
  • MxGPS achieves this at 1.6M parameters, an order of magnitude fewer than the reference baseline, suggesting multi-task training is a parameter-efficient route to topology-agnostic generalisation.
  • The voltage-magnitude collapse seen in single-task SSE fine-tuning is prevented by joint training, indicating the benefit spans accuracy calibration, not just boundary compliance.

Where Pith is reading between the lines

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

  • The paper leaves untested whether the 0% BVR and low degradation come specifically from the joint SSE+PF gradient interaction or from the extra physics and boundary losses that only MxGPS receives; a single-task GPS trained with the same auxiliary losses would isolate this.
  • The authors expect cross-branch attention to become more useful at larger task portfolios (e.g., adding OPF or N-1 contingency); a testable prediction is that K>2 will show cross-branch attention outperforming the independent-branch configuration.
  • The observed in-distribution slope of 0.48–0.58 for MxGPS on PF suggests amplitude shrinkage; one inference is that adding a calibration or magnitude-matching term to the joint loss could close the in-distribution accuracy gap without sacrificing zero-shot robustness.
  • The sliding-window protocol could be applied to other scientific GNN domains (e.g., molecular or traffic networks) to check whether topology overfitting is a general failure mode of single-task graph fine-tuning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes MxGPS, a multiplex graph transformer with a shared node encoder and K task-specialised GPS branches, jointly trained on static state estimation (SSE) and AC power flow (PF) after MGT self-supervised pre-training. The central claim is that single-task GNN fine-tuning suffers from 'topology overfitting' — low in-distribution error coupled with large degradation under topology shift — and that joint multi-task training acts as a structural mechanism against it. The reported evidence is that MxGPS achieves 0% boundary violation rate on all four zero-shot PF topologies and degrades by only 39% in PF MAE, while single-task baselines with lower in-distribution error degrade by 190%–1400%.

Significance. If the central claim holds, the paper makes a useful conceptual and practical contribution: it identifies a failure mode (topology overfitting) that is not captured by in-distribution accuracy, and proposes a parameter-efficient architecture that improves zero-shot topology transfer through multi-task training. The paper is also refreshing in its use of calibration metrics (regression slope and R²) to expose mean-collapse, and the GPS-vs-MxGPS comparison is a well-motivated controlled ablation. The empirical protocol is described in enough detail to be reproduced. However, the headline results are currently confounded by the fact that MxGPS alone receives the boundary penalty and physics loss, and the zero-shot evaluation uses a per-case normaliser fitted on a held-out topology. These issues must be resolved before the causal claim about multi-task regularization can be accepted.

major comments (4)
  1. [Section V-D and Eq. (12)] The central inference — that joint SSE+PF training, rather than the extra loss terms, causes MxGPS's 0% zero-shot BVR and low degradation — is not supported by the current experimental design. MxGPS is trained with the full objective of Eq. (12), including λ_bnd L_bnd = 0.01 and λ_phys L_phys = 0.01. All single-task baselines are trained without L_bnd, and for GCN/GAT/GPS/GNS the L_phys term is zero. Since L_bnd directly penalises the reported metric (BVR), the 0% BVR could be a direct effect of the boundary penalty. Similarly, the physics loss may shrink predictions toward feasible regions. The rebuttal about clipping (Section VI-D) addresses amplitude shrinkage, not this loss confound. To support the claimed mechanism, the authors should train a single-task GPS with the same λ_bnd and λ_phys (and ideally also a MxGPS variant without L_bnd/L_phys). If that single-task GPS also attains 0
  2. [Section V-A, Table IV] The zero-shot evaluation protocol fits a per-case normaliser on 'a single representative scenario from that case' drawn from the held-out topology. This introduces information from the zero-shot case into the evaluation pipeline. If the normaliser only infers the nominal-voltage constant, the leakage is mild, but the text also says 'a per-case normaliser is fitted', which is ambiguous. The authors should state exactly which statistics are computed from the representative scenario and provide a sensitivity check using a global normaliser or a normaliser fitted only on training cases. Without this, the claim that these are purely zero-shot topologies is weaker than stated.
  3. [Section V-D, Table V] All headline numbers come from a single fixed seed ('Seed is fixed for reproducibility', Section V-D). The degradation ratios +39% vs +190%–1400% and the 0% BVR values are reported without variance estimates. Given that the central claim depends on the magnitude ordering of these ratios, the authors should report multi-seed means and standard deviations (or at least a small number of seeds) for the key Tables II–V. The Limitations section (Section VII) lists multi-seed replication as future work, but it is load-bearing for the paper's main conclusion, not an optional refinement.
  4. [Section VI-D, Table IV] The zero-shot advantage of MxGPS is not uniform across topologies: on the 14- and 24-bus cases several baselines have lower MAE Vm, and on the 300-bus case GNS-kit has lower MAE. The paper's emphasis on '0% BVR on all four' is weakened by the fact that this is the exact metric MxGPS's loss directly optimises. The discussion does note the concentration of the advantage on larger systems, but the abstract and conclusion state the result more categorically. The causal story should be tempered unless the loss confound is addressed.
minor comments (4)
  1. [Abstract] Typo: 'thistopology' should be 'this topology'.
  2. [Section V-D] The sentence 'PF uses 0.90·MSE + 0.10·L_phys, where the physics term is a per-layer residual exposed only by GNS-kit's architecture' is confusing: if the term is zero for GCN/GAT/GPS/GNS, then the effective loss for those models is pure MSE. This should be clarified to avoid implying the baselines received a physics loss.
  3. [Eq. (13)] The SCE loss notation is ambiguous: the numerator mixes \hat{x}_{if} and x_{if} without clear indexing, and the exponent γ=2.0 is written as a superscript on the whole expression. Please rewrite with explicit hats and parentheses.
  4. [Acknowledgments] 'necessary' should be 'necessarily'.

Circularity Check

1 steps flagged

Headline 0% BVR is partly trained-in via L_bnd, not an independent confirmation of multi-task mechanism.

specific steps
  1. fitted input called prediction [Section V-C (BVR metric), Section V-D (training losses), Eq. (16), Table IV]
    "BVR (%): Boundary Violation Rate, defined as the fraction of bus predictions where Vˆm,i /∈ [Vmin m,i, Vmax m,i], measuring operational safety; ... MxGPS is the only model trained with the full objective of Eq. (12), using per-task weights λk = 1.0 (SSE and PF), λphys = 0.01, and λbnd = 0.01 ... no boundary penalty Lbnd is applied to any single-task baseline."

    The reported zero-shot BVR is the exact quantity that MxGPS's boundary loss L_bnd (Eq. 16) is designed to minimize: L_bnd is a squared-ReLU penalty on Vm outside [Vmin,Vmax], while BVR counts the fraction of such violations. MxGPS alone is trained with λ_bnd=0.01; GCN/GAT/GPS/GNS/GNS-kit are not. Consequently, MxGPS's 0% BVR on the four zero-shot topologies can be produced by the boundary penalty (or by the amplitude shrinkage it encourages—MxGPS's ID PF slope is 0.48–0.58, and the Naive Vm=1 baseline also reaches 0% BVR), without invoking the joint SSE+PF gradient interaction. The paper nevertheless uses this 0% BVR as the strongest evidence that multi-task joint training is the mechanism against topology overfitting. Because the GPS-vs-MxGPS comparison simultaneously changes multi-task t

full rationale

This is primarily an empirical paper rather than a derivation, so the extreme circularity patterns (self-definitional claims, imported uniqueness theorems, ansatz-by-self-citation, renaming known results) do not apply. The zero-shot topologies are genuinely held out, and the paper reports absolute errors and calibration metrics alongside relative degradation, which weakens any charge that the results are forced by construction. However, the central evidence for the claimed mechanism—0% BVR on all four zero-shot PF topologies—is confounded: MxGPS is the only model trained with L_bnd, a soft penalty on the exact metric being reported, while all single-task baselines omit it. The paper's rebuttal to the amplitude-shrinkage objection (post-hoc clipping cannot explain low absolute error) does not rule out the loss confound, and the Limitations section does not flag the missing single-task-with-identical-losses control. Thus the strongest headline result is partially built into the training objective, though the 39% PF-MAE degradation and external zero-shot evaluation retain independent content. No load-bearing self-citation chain is present. Score 4 reflects this partial, non-derivational circularity rather than full by-construction equivalence.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

The central claims rest on a handful of hand-chosen loss weights, one conceptual entity, and domain assumptions about solver-generated labels and the physics loss. The most consequential free parameter is λ_bnd, which directly targets the BVR metric that headlines the paper.

free parameters (5)
  • boundary penalty weight λ_bnd = 0.01
    Eq. (12); only MxGPS is trained with this term. Directly penalizes predictions outside voltage limits, so it shapes the BVR metric that anchors the paper's main claim.
  • physics loss weight λ_phys = 0.01
    Eq. (12)/(15); applied to MxGPS branches during fine-tuning. Single-task baselines do not receive it, making the controlled comparison with GPS not controlled on this axis.
  • task weights λ_k = 1.0 for SSE and PF
    Eq. (12); chosen by hand; balances the two supervised losses.
  • dimension weights w_f = [0.10, 0.10, 0.35, 0.35] for Pg, Qg, Vm, Va
    Eq. (14); hand-set weights used for both SSE and PF losses.
  • SSE corruption hyperparameters = σ=0.02 p.u., ρ_mask=0.2
    Section II(c); noise and masking levels for the stress-test SSE task; authors state they are not calibrated to a metering model.
axioms (4)
  • domain assumption Newton-Raphson AC power flow solutions generated by gridfm-datakit are ground-truth physically valid operating points.
    All supervised labels and the BIM physics check are derived from solver outputs; any dataset bias or convergence artifact propagates into every accuracy and BVR number. Section V-A.
  • domain assumption The Bus Injection Model (Eqs. 2–3) with the dataset's admittance features is an adequate physical constraint.
    The physics loss L_phys assumes predicted voltages can be plugged into BIM using per-case admittance without modeling error. Section III-F.
  • domain assumption GraphGPS/GatedGCN/Transformer architectures can represent the PF and SSE mappings at the tested scale.
    No expressivity or convergence result is given; the paper relies on empirical performance to justify the architecture class. Section III-C.
  • ad hoc to paper The zero-shot per-case normalizer can be fitted on one representative scenario from the held-out case.
    Section V-A; uses a test-distribution statistic to set normalization constants, slightly weakening the 'zero-shot' purity of the evaluation.
invented entities (1)
  • Topology overfitting no independent evidence
    purpose: Explanatory construct for the observed inversion between low in-distribution error and large zero-shot degradation in single-task GNNs.
    The name is a post-hoc interpretation of the paper's own results; the controlled test that would isolate it (single-task GPS trained with the same losses) is absent, so it has no independent falsifiable handle beyond the reported pattern. Sections I and V-D.

pith-pipeline@v1.3.0-alltime-deepseek · 16064 in / 14239 out tokens · 136791 ms · 2026-08-02T03:45:59.150635+00:00 · methodology

0 comments
read the original abstract

Single-task fine-tuning of graph neural networks (GNNs) for power grid problems exhibits a systematic failure mode: models that achieve the lowest in-distribution error degrade the most under topology shift. We term this topology overfitting: the tendency of task-specific gradient signals to encode relational structure particular to the training topologies rather than the underlying physics, causing models to fail on unseen grids despite strong in-distribution performance. To expose and address this failure mode, we introduce MxGPS (Multiplex GPS), a multiplex graph transformer that runs K task-specialised GPS branches over a shared node encoder, jointly trained on Static State Estimation (SSE) and AC Power Flow (PF) via a self-supervised pre-training and multi-task fine-tuning protocol, with a cross-branch attention module evaluated in ablation. The joint SSE+PF objective forces the shared encoder to simultaneously satisfy complementary gradient signals, preventing it from overfitting to topology-specific relational structure. Under a 3-fold sliding-window cross-validation spanning four unseen topologies (14-, 24-, 162-, and 300-bus), MxGPS attains 0% boundary violation rate (BVR) on all four zero-shot Power Flow topologies. Critically, models with substantially lower in-distribution PF error degrade by 190% to 1400% under topology shift, whereas MxGPS degrades by only 39%, an inversion that directly implicates topology overfitting as the failure mechanism rather than insufficient model capacity. With only 1.6M parameters (12x fewer than the GridFM reference baseline), MxGPS demonstrates that multi-task joint training is a principled and parameter-efficient mechanism for topology-agnostic generalisation in power grid foundation models.

Figures

Figures reproduced from arXiv: 2607.13763 by Antonello Monti, Arsam Aryandoust, Charilaos Papaioannou, Dimitris Giannakakos, Elissaios Sarmas, Ioannis Tsantilas, Jochen Cremer, Perdo P. Vergara, Ricardo J. Bessa, Sotiris Pelekis, Vangelis Marinakis, Vasilis Michalakopoulos.

Figure 1
Figure 1. Figure 1: MxGPS architecture with K=2 branches. A shared node encoder maps raw bus features xv and structural positional encodings (RWSE, LapPE) to a common embedding hv, which is broadcast to both task-specialised GPS branches. Each branch processes the identical physical graph topology through its own GPS layers, conditioned by a learnable task token tk. Dashed vertical lines indicate the optional cross-branch att… view at source ↗
Figure 2
Figure 2. Figure 2: SSE zero-shot MAEVm across the four unseen topologies. Among models with meaningful in-distribution SSE accuracy (GNS, GNS-kit, MxGPS), MxGPS variants degrade the least (Table V); GPS, GAT, and GCN show artificially low degradation due to in-distribution magnitude collapse (Table III). VI. RESULTS AND DISCUSSION A. Masked Grid Task Pre-training Table I serves one purpose: to show that MGT pre-training carr… view at source ↗
Figure 3
Figure 3. Figure 3: In-distribution boundary violation rate per model and task. MxGPS variants consistently achieve the lowest violation rates, particularly on PF. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Zero-shot PF accuracy–efficiency Pareto frontier: average MAE [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 5 linked inside Pith

  1. [1]

    Review of load-flow calculation methods,

    B. Stott, “Review of load-flow calculation methods,”Proceedings of the IEEE, vol. 62, no. 7, pp. 916–929, 1974

    1974

  2. [2]

    A review of graph neural networks and their applications in power systems,

    W. Liao, B. Bak-Jensen, J. R. Pillai, Y . Wang, and Y . Wang, “A review of graph neural networks and their applications in power systems,”Journal of Modern Power Systems and Clean Energy, vol. 10, no. 2, pp. 345– 360, 2021

    2021

  3. [3]

    Safepowergraph: Safety- aware evaluation of graph neural networks for transmission power grids,

    S. Ghamizi, A. Bojchevski, A. Ma, and J. Cao, “Safepowergraph: Safety- aware evaluation of graph neural networks for transmission power grids,” arXiv preprint arXiv:2407.12421, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Graph neural networks: A review of methods and applications,

    J. Zhou, G. Cui, S. Hu, Z. Zhang, C. Yang, Z. Liu, L. Wang, C. Li, and M. Sun, “Graph neural networks: A review of methods and applications,” AI open, vol. 1, pp. 57–81, 2020

    2020

  5. [5]

    Deepopf+: A deep neural network approach for dc optimal power flow for ensuring feasibility,

    T. Zhao, X. Pan, M. Chen, A. Venzke, and S. H. Low, “Deepopf+: A deep neural network approach for dc optimal power flow for ensuring feasibility,” in2020 IEEE international conference on communications, control, and computing technologies for smart grids (SmartGridComm), pp. 1–6, IEEE, 2020

    2020

  6. [6]

    Foundation models for the electric power grid,

    H. F. Hamann, B. Gjorgiev, T. Brunschwiler, L. S. Martins, A. Puech, A. Varbella, J. Weiss, J. Bernabe-Moreno, A. B. Mass ´e, S. L. Choi, et al., “Foundation models for the electric power grid,”Joule, vol. 8, no. 12, pp. 3245–3258, 2024

    2024

  7. [7]

    Recipe for a general, powerful, scalable graph transformer,

    L. Ramp ´aˇsek, M. Galkin, V . P. Dwivedi, A. T. Luu, G. Wolf, and D. Beaini, “Recipe for a general, powerful, scalable graph transformer,” vol. 35, pp. 14501–14515, 2022

    2022

  8. [8]

    Masked au- toencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked au- toencoders are scalable vision learners,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 16000– 16009, 2022

    2022

  9. [9]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” inPro- ceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pp. 4171–4186, 2019

    2019

  10. [10]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in neural information processing systems, vol. 30, 2017

    2017

  11. [11]

    Moe-graphsage-based integrated evaluation of transient rotor angle and voltage stability in power systems,

    K. Zhang, G. Yang, F. Shi, S. He, and Y . Zhang, “Moe-graphsage-based integrated evaluation of transient rotor angle and voltage stability in power systems,”arXiv preprint arXiv:2511.08610, 2025

    arXiv 2025

  12. [12]

    Attending to graph transformers,

    L. M ¨uller, M. Galkin, C. Morris, and L. Ramp ´aˇsek, “Attending to graph transformers,”arXiv preprint arXiv:2302.04181, 2023

    Pith/arXiv arXiv 2023

  13. [13]

    Residual gated graph convnets,

    X. Bresson and T. Laurent, “Residual gated graph convnets,”arXiv preprint arXiv:1711.07553, 2017

    Pith/arXiv arXiv 2017

  14. [14]

    Benchmarking graph neural networks,

    V . P. Dwivedi, C. K. Joshi, A. T. Luu, T. Laurent, Y . Bengio, and X. Bresson, “Benchmarking graph neural networks,”Journal of Machine Learning Research, vol. 24, no. 43, pp. 1–48, 2023

    2023

  15. [15]

    Graph neural networks with learnable structural and positional representations,

    V . P. Dwivedi, A. T. Luu, T. Laurent, Y . Bengio, and X. Bresson, “Graph neural networks with learnable structural and positional representations,” inThe Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, OpenReview.net, 2022

    2022

  16. [16]

    Heterogeneous graph trans- former,

    Z. Hu, Y . Dong, K. Wang, and Y . Sun, “Heterogeneous graph trans- former,” inProceedings of the web conference 2020, pp. 2704–2710, 2020

    2020

  17. [17]

    Graphmae: Self-supervised masked graph autoencoders,

    Z. Hou, X. Liu, Y . Cen, Y . Dong, H. Yang, C. Wang, and J. Tang, “Graphmae: Self-supervised masked graph autoencoders,” inProceed- ings of the 28th ACM SIGKDD conference on knowledge discovery and data mining, pp. 594–604, 2022

    2022

  18. [18]

    Multitask learning,

    R. Caruana, “Multitask learning,”Machine learning, vol. 28, no. 1, pp. 41–75, 1997

    1997

  19. [19]

    Multi-task graph neural architecture search with task-aware collaboration and curriculum,

    Y . Qin, X. Wang, Z. Zhang, H. Chen, and W. Zhu, “Multi-task graph neural architecture search with task-aware collaboration and curriculum,” Advances in neural information processing systems, vol. 36, pp. 24879– 24891, 2023

    2023

  20. [20]

    Graph mixture of experts: Learning on large-scale graphs with explicit diversity modeling,

    H. Wang, Z. Jiang, Y . You, Y . Han, G. Liu, J. Srinivasa, R. Kompella, Z. Wang,et al., “Graph mixture of experts: Learning on large-scale graphs with explicit diversity modeling,”Advances in neural information processing systems, vol. 36, pp. 50825–50837, 2023

    2023

  21. [21]

    Mind the links: Cross-layer attention for link prediction in multiplex networks,

    D. Sharma, A. Kishore, A. Garg, D. Mazumder, D. Mohapatra, and J. Patro, “Mind the links: Cross-layer attention for link prediction in multiplex networks,” inProceedings of the Nineteenth ACM Interna- tional Conference on Web Search and Data Mining, pp. 1243–1247, 2026

    2026

  22. [22]

    Physics-informed machine learning,

    G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang, “Physics-informed machine learning,”Nature Reviews Physics, vol. 3, no. 6, pp. 422–440, 2021

    2021

  23. [23]

    Language mod- els are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell,et al., “Language mod- els are few-shot learners,”Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020

    1901

  24. [24]

    gridfm-datakit-v1: A python library for scalable and realistic power flow and optimal power flow data generation,

    A. Puech, M. Mazzonelli, C. Cintas, T. R. Govindasamy, M. Mn- gomezulu, J. Weiss, M. Ba `u, A. Varbella, F. Mirall `es, K. Kim, et al., “gridfm-datakit-v1: A python library for scalable and realistic power flow and optimal power flow data generation,”arXiv preprint arXiv:2512.14658, 2025

    arXiv 2025

  25. [25]

    Mat- power: Steady-state operations, planning, and analysis tools for power systems research and education,

    R. D. Zimmerman, C. E. Murillo-S ´anchez, and R. J. Thomas, “Mat- power: Steady-state operations, planning, and analysis tools for power systems research and education,”IEEE Transactions on power systems, vol. 26, no. 1, pp. 12–19, 2010

    2010

  26. [26]

    The power grid library for benchmarking ac optimal power flow algorithms,

    S. Babaeinejadsarookolaee, A. Birchfield, R. D. Christie, C. Coffrin, C. DeMarco, R. Diao, M. Ferris, S. Fliscounakis, S. Greene, R. Huang, et al., “The power grid library for benchmarking ac optimal power flow algorithms,”arXiv preprint arXiv:1908.02788, 2019

    Pith/arXiv arXiv 1908

  27. [27]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” in5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Con- ference Track Proceedings, OpenReview.net, 2017

    2017

  28. [28]

    Graph attention networks,

    P. Velickovic, G. Cucurull, A. Casanova, A. Romero, P. Li `o, and Y . Bengio, “Graph attention networks,” in6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings, OpenReview.net, 2018

    2018

  29. [29]

    How attentive are graph attention net- works?,

    S. Brody, U. Alon, and E. Yahav, “How attentive are graph attention net- works?,” inThe Tenth International Conference on Learning Represen- tations, ICLR 2022, Virtual Event, April 25-29, 2022, OpenReview.net, 2022

    2022

  30. [30]

    Gridsfm: A foundation model for ac optimal power flow

    W. Yang, A. Britto Mattos Lima, T. V . Spina, S. Fowers, B. Zhang, and C. White, “Gridsfm: A foundation model for ac optimal power flow.” May 2026

    2026

  31. [31]

    Fast graph representation learning with pytorch geometric,

    M. Fey and J. E. Lenssen, “Fast graph representation learning with pytorch geometric,”arXiv preprint arXiv:1903.02428, 2019

    Pith/arXiv arXiv 1903

  32. [32]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” in 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019, OpenReview.net, 2019

    2019