Pith. sign in

REVIEW 4 major objections 5 minor 33 references

Generative OPF becomes feasible and scalable by separating compression from generation with latent flow matching.

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

2026-08-01 05:17 UTC pith:TU7RCKWS

load-bearing objection A sensible first application of flow matching to AC-OPF, undermined by internal inconsistencies in the feasibility claims and a confounded ablation; worth referee time but needs major revision. the 4 major comments →

arxiv 2607.22788 v1 pith:TU7RCKWS submitted 2026-07-24 cs.LG cs.AIstat.ML

FMOPF: Latent Flow Matching with Constraint-Aware Interaction Priors for AC Optimal Power Flow

Zhilin Huang This is my paper
classification cs.LG cs.AIstat.ML
keywords optimal power flowgenerative modelingflow matchinglatent spaceconstraint-aware priorfeasibilitytail riskpower systems
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 argues that generative models for AC optimal power flow fail because they try to do two jobs at once: compress the high-dimensional set of valid dispatches and learn how dispatches depend on load. FMOPF separates these jobs with a two-stage pipeline: an autoencoder learns a compact latent space for dispatches, and a flow-matching model samples that space conditioned on loads, with a constraint-aware interaction prior that injects load-state coupling and power-flow residual corrections. The paper claims this decoupling is a necessary condition for generating physically feasible, diverse dispatches, and is what allows the method to be the first generative OPF approach that works at 300-bus scale with full feasibility and the lowest worst-case cost among generative methods.

Core claim

On its own terms, FMOPF's central discovery is that a learned latent representation of the feasible OPF solution manifold, combined with flow matching in that latent space, yields generations that are inherently biased toward feasibility, whereas diffusion models that operate directly in the raw joint load-state space conflate compression with conditional generation and degrade. The constraint-aware interaction prior computes a multiplicative coupling between load and state features at each generation step, then gates in a correction supervised by power-flow residuals; ablations show this prior acts as a late-stage tail-risk controller, and removing it raises worst-case cost by orders of mag

What carries the argument

The key machinery is the latent flow-matching pipeline. Stage one trains an autoencoder with a mean-squared reconstruction loss to map each 4n-dimensional OPF state to a latent vector and back. Stage two freezes the autoencoder and trains, jointly, a flow-matching velocity field v_theta(z_t,t,l,f) and the CA-IPN: an MLP pair encodes loads and the current decoded state, combines them by element-wise product to form a cross feature, and fuses this with a constraint-correction feature through a learned gate. The velocity field interpolates linearly between data latent and Gaussian noise; at inference, Euler integration in the latent space produces a latent, which is decoded and refined by Newto

Load-bearing premise

The load-bearing premise is that training an autoencoder with a reconstruction loss forces the latent space to retain all physical variables and makes decoded states inherently biased toward the feasible region, even when the latent is larger than the state itself.

What would settle it

On the 24-bus system, scale the joint-space variant's hidden width and training epochs to match FMOPF's total compute; if its post-refinement box feasibility rises from roughly 45% toward 100%, the advocated necessity of latent compression is not established. A second check: decode random latent vectors from the trained autoencoder and count how many are box-feasible; if random latents already decode to feasible states, the feasibility bias is a decoder artifact rather than a property of learned compression.

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

If this is right

  • If the central claim holds, joint-space diffusion for OPF is not just weaker but conceptually mistargeted; future generative OPF work should operate in a learned latent space.
  • Operators can use FMOPF's generated samples for risk quantification and sensitivity analysis at scales where previous generative methods produced mostly infeasible samples.
  • Because FMOPF consistently lands inside the feasible region before refinement, it provides Newton-Raphson warm starts that converge in one to three iterations, cutting computation per solve.
  • The design identifies autoencoder compression ratio as the primary scalability lever, pointing toward hierarchical latents for multi-thousand-bus systems.
  • The method's controlled cost diversity separates useful near-optimal exploration from the inflated variance of degraded joint-space samples.

Where Pith is reading between the lines

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

  • The paper leaves implicit that the same decoupling should transfer to other constrained optimization sampling tasks, such as security-constrained OPF, unit commitment, or gas networks; if compression is genuinely the bottleneck, any task with a low-dimensional feasible manifold could benefit.
  • A testable extension would be to replace the learned autoencoder with a physics-based compression, such as PCA on the solution manifold or a decoder that enforces power-flow structure, and measure whether the feasibility bias persists; if it disappears, the bias is not inherent to latent spaces per se but to how the encoder is trained.
  • The ablation's 'late 30%' timing result suggests the generation trajectory has a coarse-to-fine structure; an adaptive step-size or two-phase sampler with early unconditional generation and late conditioned refinement might cut compute further.
  • On the 6-bus system the latent dimension (64) exceeds the state dimension (24), so no compression occurs; the paper's necessity claim could be sharpened by testing latent dimensions below versus above the manifold's intrinsic dimension.

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 / 5 minor

Summary. The paper proposes FMOPF, a two-stage generative framework for AC optimal power flow. Stage 1 trains an autoencoder to compress 4n-dimensional OPF states into a latent space; Stage 2 trains a conditional flow-matching model together with a Constraint-Aware Interaction Prior Network (CA-IPN) that injects load-state interaction features and constraint-residual corrections during sampling. The authors evaluate on IEEE 6/24/118/300-bus systems and claim that FMOPF provides the best Newton-Raphson warm starts, the lowest tail risk among generative baselines, and the first generative OPF method that scales to 300 buses while preserving full feasibility. The central mechanistic claim is that decoupling compression from generation through a latent autoencoder is a necessary condition for physical feasibility. I assess the evidence for that claim below.

Significance. Conditional generative modeling of near-optimal AC-OPF solutions is a timely and important problem, and the paper contains useful ingredients: a sensible two-stage design, a flow-matching formulation in a learned latent space, and an ablation in Table V that does support the CA-IPN as a tail-risk-reducing component. The 300-bus generative evaluation is more extensive than much prior work on generative OPF. However, the main conceptual claim that the latent autoencoder is necessary is not isolated by the experiments, and the numerical tables contradict the 'full feasibility' claim. The paper ships no code or data, which further limits verification. Until these load-bearing issues are resolved, the central contribution is not established.

major comments (4)
  1. [Table III and §IV.B] The abstract and §IV.B claim that FMOPF is 'the only generative method that scales to the 300-bus system while maintaining... full feasibility.' Table III directly contradicts this. After Newton-Raphson refinement, FMOPF has line feasibility 41.8±0.1% on the 6-bus system and box feasibility 98.7±0.2% on the 118-bus system. Table II also shows raw line feasibility of 41.8% on 6-bus. Thus 'full feasibility' is not supported for all systems, and the wording needs to be qualified to specify which constraints are satisfied where. This is not a minor wording issue: the feasibility claim is a central advertised result.
  2. [§III-C vs. Table I, Table IV, §IV-E] The latent dimension is reported inconsistently. §III-C states d_z = 64 for the 6- and 24-bus systems, 128 for the 118-bus system, and 256 for the 300-bus system. Table I's footnote says the compression ratio uses d_z = 256 for all systems; Table IV states FMOPF uses a 256-dimensional latent space; and §IV-E says the flow matching model and CA-IPN operate in a fixed 256-dimensional latent space. These cannot all be correct. Moreover, on the 6-bus system the state dimension is 4n = 24, so a 64- or 256-dimensional latent is an expansion, not a compression (Table I reports 0.09×). The 'compression' mechanism that the paper credits for FMOPF's success is therefore absent on the smallest systems, and the reported numbers do not support the scalability/compression narrative.
  3. [§IV-C, Table IV] The ablation that supports the central claim is confounded. Table IV compares FMOPF (latent space + flow matching) against 'w/o Latent' (joint raw state space + DDIM sampling). This varies two factors simultaneously: the representation space and the generative sampler. The text itself says the joint-space variant 'enforces exact load matching' via DDIM hard measurement replacement, so the observed feasibility collapse could be an artifact of DDIM's conditioning mechanism rather than evidence that a latent space is necessary. To support the claim that the latent autoencoder is the cause, the authors need a controlled comparison that changes only the representation (e.g., latent flow matching vs. joint-space flow matching, or latent DDIM vs. joint-space DDIM), and should report the full metric set, including line and box feasibility and power-flow error, not only optimality gap and box fea
  4. [§IV-C] The assertion that 'the reconstruction objective forces the latent space to retain information about all physical variables, and states decoded from this space are inherently biased toward the feasible region' is unsupported and appears contradicted by the paper's own results. An autoencoder trained with an MSE reconstruction loss on feasible OPF solutions can memorize the training distribution without providing any guarantee that decoded states satisfy nonlinear AC power-flow equations or operational limits. Tables II and III show FMOPF line feasibility of only 41.8% on the 6-bus system and box feasibility of 98.7% on the 118-bus system after refinement, so the decoded states are not demonstrably 'inherently biased' toward the full feasible region. If the authors intend this as a theoretical statement, a proof or a carefully designed empirical test is needed; as written, it is an assert
minor comments (5)
  1. [Reproducibility] No code, data, or trained model checkpoints are provided. Given that the paper relies on five-seed means and several hyperparameters that are not fully specified, releasing the evaluation pipeline would substantially improve verifiability.
  2. [Eq. (9)] The symbol t is used both for the flow-matching time variable and for the constraint-residual target in Eq. (9). This overloading is confusing; rename one of them.
  3. [Tables II/III and §IV.B] In §IV.B the text says FMOPF 'maintains full box feasibility before any post-processing,' but Table III reports box feasibility 98.7% on the 118-bus system after refinement. Please clarify whether the 'before post-processing' statement refers only to raw outputs and why the refined value drops below 100%.
  4. [Table VII] The table lists 'FMOPF w/o Latent' with parameters growing to 82.0M at 300 buses. This appears to be the joint-space DDIM variant from Table IV, but the connection is not stated explicitly. Clarify so a reader can map the ablation configuration to the complexity numbers.
  5. [Table VI] The diversity experiment is reported with cost standard deviations in 'pu' across systems of very different absolute cost scales. A normalized measure or the underlying cost magnitudes would help interpret whether the reported diversity is economically meaningful.

Circularity Check

0 steps flagged

No significant circularity: central claims are empirical benchmark comparisons against external baselines and solvers; the only self-citation is transparent and not load-bearing.

full rationale

FMOPF's main claims—warm-start quality, tail risk, scalability, and feasibility—are evaluated as benchmark outcomes after Newton-Raphson refinement against independent baselines and pandapower ground truth, so no prediction is derived from a fitted parameter or defined into existence. The CA-IPN is adapted from the author's IPDiff paper [31], but that citation is used as architectural inspiration in related work ('Our CA-IPN extends this concept to the OPF domain'), not as an unverified uniqueness theorem or as the proof of the paper's empirical results. The assertion that 'states decoded from this space are inherently biased toward the feasible region' is not proved and the Table IV ablation is confounded by changing both the representation space and the sampler; however, a confounded ablation is a causal-inference weakness, not a circular derivation. The latent-dimension inconsistency (Section III.C gives d_z=64/128/256, while Table I uses d_z=256, yielding ratios 0.09x/0.38x on small systems) is an internal-consistency issue, not an equation that reduces to its input. No step exhibits the required form: an output equivalent by construction to a fitted input, a definition in terms of the predicted quantity, or a load-bearing argument resting solely on a self-citation. Score 1 rather than 0 only to acknowledge the transparent self-citation to IPDiff; it does not affect the derivation.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 1 invented entities

The central claims rest almost entirely on empirical benchmark results. The listed free parameters are hand-chosen hyperparameters; the key domain assumption is that the autoencoder's reconstruction loss biases decoded states toward the feasible region. No new physical entities are introduced; CA-IPN is an architectural invention evaluated only on the paper's own benchmarks.

free parameters (7)
  • Latent dimension d_z = 64 (6/24-bus), 128 (118-bus), 256 (300-bus)
    Hand-selected; controls compression ratio and underpins the scaling claim. Table I says fixed 256, contradicting §III-C.
  • Flow-matching sampling steps K = 30
    Chosen without sensitivity analysis; the efficient-sampling claim depends on K being adequate.
  • Constraint loss weight λ_con = 0.1
    Hand-tuned balance in Eq. (8); tail-risk behavior depends on it.
  • Coarse loss weight λ_c = 0.3
    Hand-tuned; the coarse pass is used to compute the CA-IPN input.
  • Constraint correction scale = 0.01
    Residuals are scaled before supervision; tanh bounds the correction to [-1,1], limiting the physics-based correction.
  • Load perturbation range = ±15%
    Dataset-generating distribution; generalization beyond this range is untested.
  • Architecture hidden sizes = AE 1024, flow 4096, CA-IPN 1024
    Fixed hyperparameters; no ablation on model capacity.
axioms (5)
  • domain assumption AC-OPF ground truth from pandapower on IEEE test systems is representative of real grid operations.
    All results depend on this; no real-network validation is provided.
  • ad hoc to paper Autoencoder reconstruction loss biases decoded states toward the feasible OPF manifold.
    Asserted in §IV.C; no proof or metric such as decoded-feasibility rate before refinement is reported.
  • domain assumption Euler ODE integration with 30 steps yields valid samples from the conditional latent distribution.
    No convergence checks or step-count sensitivity analysis is provided.
  • domain assumption Newton-Raphson with analytical Jacobian converges in 1-3 iterations from FMOPF warm starts.
    Used to guarantee post-refinement feasibility; baselines show divergence, so warm-start quality is doing the work.
  • ad hoc to paper The 'conflation' of compression and generation is the root cause of failure in prior methods.
    Motivating hypothesis; the ablation changes multiple variables at once and does not isolate conflation cleanly.
invented entities (1)
  • Constraint-Aware Interaction Prior Network (CA-IPN) no independent evidence
    purpose: Models load-state coupling and injects constraint-aware corrections at each flow-matching step; claimed late-stage tail-risk controller.
    The CA-IPN is evaluated only on the paper's own benchmark tables and ablations. No external falsifiable handle is provided, and its main design is adapted from the author's IPDiff.

pith-pipeline@v1.3.0-alltime-deepseek · 16670 in / 13187 out tokens · 107532 ms · 2026-08-01T05:17:27.137360+00:00 · methodology

0 comments
read the original abstract

AC optimal power flow determines the minimum-cost generation dispatch under nonlinear power balance constraints and is solved thousands of times daily in electricity market operations. Learning a direct mapping from load conditions to OPF solutions can accelerate this computation, yet with deepening renewable penetration, a single optimal dispatch is no longer sufficient. Operators require a characterization of the distribution of feasible near-optimal solutions for risk quantification, sensitivity analysis, and multi-objective trade-off assessment. Supervised neural networks provide fast point predictions but cannot capture this conditional distribution. Diffusion-based generative models can sample diverse solutions in principle, yet existing methods operating in the raw state space exhibit degraded solution quality and fail to scale beyond medium-sized systems. We identify the root cause as the conflation of two distinct tasks within a single model. Compressing the high-dimensional OPF solution manifold is one task, and learning the conditional mapping from loads to that manifold is another. This paper presents FMOPF, a framework that resolves this conflation by decoupling compression from generation through latent flow matching and by explicitly modeling load-state coupling through a Constraint-Aware Interaction Prior Network. Experiments on four IEEE test systems demonstrate that FMOPF provides the most effective Newton-Raphson warm starts, achieves the lowest tail risk among generative methods, and is the first such method to scale to systems with several hundred buses while preserving full feasibility. Ablation studies confirm that the latent generation pipeline is a necessary condition for physical feasibility and that the interaction prior functions as a late-stage tail-risk controller.

Figures

Figures reproduced from arXiv: 2607.22788 by Zhilin Huang.

Figure 1
Figure 1. Figure 1: Overview of the FMOPF framework. Stage 1 trains an autoencoder to compress OPF states into a compact latent space. [PITH_FULL_IMAGE:figures/full_fig_p003_1.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

33 extracted references · 4 linked inside Pith

  1. [1]

    An introduction to optimal power flow: Theory, formulation, and examples,

    S. Frank and S. Rebennack, “An introduction to optimal power flow: Theory, formulation, and examples,”IIE transactions, vol. 48, no. 12, pp. 1172–1197, 2016

    2016

  2. [2]

    A. J. Wood, B. F. Wollenberg, and G. B. Shebl ´e,Power generation, operation, and control. John wiley & sons, 2013

    2013

  3. [3]

    Deepopf: A deep neural network approach for security-constrained dc optimal power flow,

    X. Pan, T. Zhao, M. Chen, and S. Zhang, “Deepopf: A deep neural network approach for security-constrained dc optimal power flow,”IEEE Transactions on Power Systems, vol. 36, no. 3, pp. 1725–1735, 2020

    2020

  4. [4]

    Learning optimal solutions for extremely fast ac optimal power flow,

    A. S. Zamzam and K. Baker, “Learning optimal solutions for extremely fast ac optimal power flow,” in2020 IEEE international conference on communications, control, and computing technologies for smart grids (SmartGridComm). IEEE, 2020, pp. 1–6

    2020

  5. [5]

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

    F. Fioretto, T. W. Mak, and P. Van Hentenryck, “Predicting ac optimal power flows: Combining deep learning and lagrangian dual methods,” inProceedings of the AAAI conference on artificial intelligence, vol. 34, no. 01, 2020, pp. 630–637

    2020

  6. [6]

    Model-informed generative adversarial network for learning optimal power flow,

    Y . Li, C. Zhao, and C. Liu, “Model-informed generative adversarial network for learning optimal power flow,”IISE Transactions, vol. 57, no. 1, pp. 30–43, 2025

    2025

  7. [7]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems, vol. 33, pp. 6840– 6851, 2020

    2020

  8. [8]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inThe Eleventh International Con- ference on Learning Representations, 2023

    2023

  9. [9]

    Diffusion-based scene graph to im- age generation with masked contrastive pre-training,

    L. Yang, Z. Huang, Y . Song, S. Hong, G. Li, W. Zhang, B. Cui, B. Ghanem, and M.-H. Yang, “Diffusion-based scene graph to im- age generation with masked contrastive pre-training,”arXiv preprint arXiv:2211.11138, 2022

    Pith/arXiv arXiv 2022

  10. [10]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

    2022

  11. [11]

    Scaling rectified flow transformers for high-resolution image synthesis,

    P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. M ¨uller, H. Saini, Y . Levi, D. Lorenz, A. Sauer, F. Boeselet al., “Scaling rectified flow transformers for high-resolution image synthesis,” inForty-first international conference on machine learning, 2024

    2024

  12. [12]

    Motion-aware latent diffusion models for video frame interpolation,

    Z. Huang, Y . Yu, L. Yang, C. Qin, B. Zheng, X. Zheng, Z. Zhou, Y . Wang, and W. Yang, “Motion-aware latent diffusion models for video frame interpolation,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 1043–1052. 11

    2024

  13. [13]

    Enhanced motion-aware latent diffusion models for video frame interpolation,

    Z. Huang, C. Qin, Y . Xing, and W. Yang, “Enhanced motion-aware latent diffusion models for video frame interpolation,” inProceedings of the 33rd ACM International Conference on Multimedia, 2025, pp. 2811–2820

    2025

  14. [15]

    Interaction-based retrieval- augmented diffusion models for protein-specific 3d molecule gener- ation,

    Z. Huang, L. Yang, X. Zhou, C. Qin, Y . Yu, X. Zheng, Z. Zhou, W. Zhang, Y . Wang, and W. Yang, “Interaction-based retrieval- augmented diffusion models for protein-specific 3d molecule gener- ation,” inForty-first International Conference on Machine Learning, 2024

    2024

  15. [16]

    Binding-adaptive diffusion models for structure- based drug design,

    Z. Huang, L. Yang, Z. Zhang, X. Zhou, Y . Bao, X. Zheng, Y . Yang, Y . Wang, and W. Yang, “Binding-adaptive diffusion models for structure- based drug design,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 11, 2024, pp. 12 671–12 679

    2024

  16. [17]

    Disentangled diffusion model for 3d molecu- lar generation with protein–ligand interaction priors,

    Z. Huang, L. Yang, C. Qin, Y . Xing, H. Yu, X. Zhou, B. Zheng, Y . Wang, X. Gao, and W. Yang, “Disentangled diffusion model for 3d molecu- lar generation with protein–ligand interaction priors,”Bioinformatics, vol. 42, no. 6, p. btag165, 2026

    2026

  17. [18]

    Diffopf: Diffusion solver for optimal power flow,

    M. Hoseinpour and V . Dvorkin, “Diffopf: Diffusion solver for optimal power flow,”Electric Power Systems Research, vol. 262, p. 113621, 2027

    2027

  18. [19]

    Fast diffusion with physics-correction for acopf,

    S. Shekhar, A. Karn, K. Keshav, S. Bansal, and P. Pareek, “Fast diffusion with physics-correction for acopf,”arXiv preprint arXiv:2602.03020, 2026

    arXiv 2026

  19. [20]

    Flow straight and fast: Learning to generate and transfer data with rectified flow,

    X. Liu, C. Gonget al., “Flow straight and fast: Learning to generate and transfer data with rectified flow,” inThe Eleventh International Conference on Learning Representations, 2023

    2023

  20. [21]

    Dc3: A learning method for optimization with hard constraints,

    P. L. Donti, D. Rolnick, and J. Z. Kolter, “Dc3: A learning method for optimization with hard constraints,”arXiv preprint arXiv:2104.12225, 2021

    Pith/arXiv arXiv 2021

  21. [22]

    Unsupervised learning for solving ac optimal power flows: Design, analysis, and experiment,

    W. Huang, M. Chen, and S. H. Low, “Unsupervised learning for solving ac optimal power flows: Design, analysis, and experiment,”IEEE Transactions on Power Systems, vol. 39, no. 6, pp. 7102–7114, 2024

    2024

  22. [23]

    Physics-informed neural net- works for ac optimal power flow,

    R. Nellikkath and S. Chatzivasileiadis, “Physics-informed neural net- works for ac optimal power flow,”Electric Power Systems Research, vol. 212, p. 108412, 2022

    2022

  23. [24]

    Topology-aware graph neural networks for learning feasible and adaptive ac-opf solutions,

    S. Liu, C. Wu, and H. Zhu, “Topology-aware graph neural networks for learning feasible and adaptive ac-opf solutions,”IEEE Transactions on Power Systems, vol. 38, no. 6, pp. 5660–5670, 2022

    2022

  24. [25]

    Deep unsupervised learning using nonequilibrium thermodynamics,

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” in International conference on machine learning. pmlr, 2015, pp. 2256– 2265

    2015

  25. [26]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” inInternational Conference on Learning Representations, 2021

    2021

  26. [27]

    Diffusion models beat gans on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,”Advances in neural information processing systems, vol. 34, pp. 8780–8794, 2021

    2021

  27. [28]

    Building normalizing flows with stochastic interpolants,

    M. S. Albergo and E. Vanden-Eijnden, “Building normalizing flows with stochastic interpolants,” inThe Eleventh International Conference on Learning Representations, 2023

    2023

  28. [29]

    Equivariant flow matching with hybrid probability transport for 3d molecule generation,

    Y . Song, J. Gong, M. Xu, Z. Cao, Y . Lan, S. Ermon, H. Zhou, and W.- Y . Ma, “Equivariant flow matching with hybrid probability transport for 3d molecule generation,”Advances in Neural Information Processing Systems, vol. 36, pp. 549–568, 2023

    2023

  29. [30]

    Se (3) diffusion model with application to protein backbone generation,

    J. Yim, B. L. Trippe, V . De Bortoli, E. Mathieu, A. Doucet, R. Barzilay, and T. Jaakkola, “Se (3) diffusion model with application to protein backbone generation,”arXiv preprint arXiv:2302.02277, 2023

    Pith/arXiv arXiv 2023

  30. [31]

    Protein-ligand interaction prior for binding-aware 3d molecule diffusion models,

    Z. Huang, L. Yang, X. Zhou, Z. Zhang, W. Zhang, X. Zheng, J. Chen, Y . Wang, B. Cui, and W. Yang, “Protein-ligand interaction prior for binding-aware 3d molecule diffusion models,” inThe Twelfth Interna- tional Conference on Learning Representations, 2024

    2024

  31. [32]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101, 2017

    Pith/arXiv arXiv 2017

  32. [33]

    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

  33. [34]

    pandapower—an open-source python tool for convenient modeling, analysis, and optimization of electric power systems,

    L. Thurner, A. Scheidler, F. Sch ¨afer, J.-H. Menke, J. Dollichon, F. Meier, S. Meinecke, and M. Braun, “pandapower—an open-source python tool for convenient modeling, analysis, and optimization of electric power systems,”IEEE Transactions on Power Systems, vol. 33, no. 6, pp. 6510– 6521, 2018

    2018