acceptodds
Under review as a conference paper at ICLR 2027

PIDSolver:A Sequential Physics-Informed Diffusion Solver for Optimization with Algebraic Constraints

Abstract

Physics-informed diffusion models have emerged as a promising tool for solving large optimization problems with hard constraints. These models iteratively generate solutions, thereby improving the accuracy of end-to-end artificial neural networks (ANN) based solvers. However, for a class of optimization problems involving complex algebraic constraints, a traditional solver is required to repeatedly solve for latent variables and perform gradient estimation for ANN-based solver training, incurring prohibitively high time costs. In this work, we present a sequential Physics-informed Diffusion Solver (PIDSolver) to address this challenge. Specifically, PIDSolver couples two sequential networks: the former performs a physics-informed denoising process, while the latter handles the algebraic constraints solving process and concurrently performs gradient-based corrections to guide the former in satisfying inequality constraints. The effectiveness of the proposed method is validated in practical alternating current optimal power flow scenarios, where PIDSolver generates solutions within a second-level time scale while outperforming state-of-the-art ANN-based solvers. Our code will be released at https://anonymous.4open.science/r/PIDSolver-D822/.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
AC optimal power flow, physics-informed neural networks, Diffusion models, optimization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_A8nFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PIDSolver, a learning-based framework for optimization with algebraic constraints, evaluated on AC optimal power flow (AC-OPF). It couples a conditional diffusion network that generates control variables with a learned power-flow surrogate that predicts the remaining physical states. Constraint residuals computed from these outputs provide end-to-end training signals, while accelerated sampling trades inference time against solution quality. The paper also presents conditional error and feasibility analyses. Experiments on the IEEE 118-bus system report lower constraint violations at comparable generation costs relative to the evaluated learning-based baselines, with additional ablation, random-seed, and load-shift experiments.

Critical strengths
  1. The two-network design is interesting and well motivated. Separating control generation from the prediction of the remaining physical states provides a coherent way to incorporate constraint feedback into diffusion training. This directly targets the computational burden of repeatedly solving power-flow equations within training and gives the framework a clear practical motivation.

  2. The reported results provide encouraging empirical evidence within the tested setting. PIDSolver reduces reactive-power violations while maintaining generation costs comparable to the diffusion baselines. The fine-tuning ablation and repeated runs across five seeds further support the usefulness and repeatability of the reported improvement on the IEEE 118-bus system.

Critical weaknesses
  1. The baseline protocols are insufficiently detailed to establish whether the comparisons evaluate complete methods under comparable conditions. In particular, it is unclear which correction, projection, or post-processing stages are retained for M1, M4, and M5, and whether their associated computation is consistently accounted for. The implementation of M2's training accelerations also needs clarification. These details could materially change the interpretation of both the constraint-violation and training-time improvements.

  2. The evaluation does not yet clearly establish the practical trade-off between speed and independently verified feasibility. Table 2 reports GPU batch times for 1,000 samples, while a corresponding comparison with conventional OPF solving under matched feasibility tolerances is missing. Furthermore, low average violations and power-flow convergence do not by themselves establish the proportion of operationally feasible outputs. Without a clearly specified independent validation procedure and comparable timing, the practical significance of the reported gains remains uncertain.

Questions for authors
  1. For M1, M4, and M5, which correction, projection, and post-processing stages were included in the comparisons, and do the reported runtimes cover those stages? Please also clarify the baseline training configurations, including whether M2 uses the acceleration mechanisms proposed in its original work.

  2. Which numerical solver and feasibility tolerances were used for the conventional OPF reference? Under the same final feasibility criteria, how does PIDSolver compare with conventional solving in single-instance latency and total solution time, including any feasibility recovery? Please distinguish these timings from the 1,000-sample batch throughput reported in Table 2.

  3. Are the reported test violations evaluated using the surrogate-predicted physical states or an independent numerical power-flow calculation? Under explicitly stated equality and inequality tolerances, what fraction of test instances is feasible, and what are the maximum or high-percentile violations? This would help determine whether the lower mean violations translate into reliable feasible solutions.

14 comments

We would like to express our sincere gratitude to the reviewer for the thorough and insightful comments, as well as for the kind recognition of our work, which are very encouraging to us. We have carefully addressed each of the concerns raised, as detailed below. In particular, the comments prompted us to fully disclose and re-measure the baseline protocols: we re-implemented M1–M5 following the original settings of each paper, stated the retained stages and the timing accounting item by item, and added matched comparisons with conventional solvers together with per-sample feasibility and tail statistics. We sincerely hope that these clarifications and revisions fully resolve the points raised, and we would be most grateful for any further comments.

Weakness 1: Regarding the baselines, the stages and runtime accounting for M1, M4, and M5 and M2’s training accelerations are insufficiently detailed.

Response. We thank the reviewer for raising this point. In response, we have carefully re-verified the reproduction of each baseline against the original settings of its paper, with all deviations explicitly noted below, and measured all methods under a single, consistent evaluation pipeline (independent PyPower verification, no clipping post-processing for any baseline or for PIDSolver).

(A) Regarding which stages are retained for M1, M4, and M5: Below, we describe the stages each method retains or omits and the implementation details (the full performance comparison and the implementation details of every method are reported in Table 1 and Appendix F). Unless otherwise noted, we adopted the hyperparameters and training protocols reported in each baseline's original paper.

M1 = a supervised deep neural network predicting (V,θ)(V,\theta), with the original method's power balance reconstruction correction step retained. Its output is used directly at evaluation time.

M2 = reproduced according to the original protocol: under the original semi-supervised protocol with all hyperparameter settings as reported in the original work. In order to compare fairly, the DNN outputs with linear activations. A differentiable Newton completion recovers remaining variables. The loss combines the supervised MSE with the weighted cost and constraint-violation terms of the original protocol.

M3 = reproduced according to the original paper, but not fully in its original unsupervised training form: if trained completely according to the original unsupervised protocol, it diverges within 2–3 epochs on all three systems, so we first performed supervised pre-training and then applied the original DC3 protocol for unsupervised training (completion stage: supervised training + per-sample numerical power-flow completion; correction stage: one Newton solve + one Jacobian inversion per sample, with gradients injected via the implicit function theorem). Even so, the correction stage still diverged within 2–3 epochs on all three systems, and applying the correction at test time to the solutions produced by M3 also diverges within 2–3 correction steps. The M3 rows in the Table 1 therefore retain the completion stage only, and their time counts only the completion computation.

M4 = the vanilla DDPM: performs stochastic reverse sampling, with no clipping and no correction of any form.

M5 = the fast constrained DDIM implemented according to "Algorithm 1" of the original paper: 30-step DDIM, η=0.2\eta=0.2, applying constraint correction x0∣t←x0∣t−λt∇(RH+RG)x_{0|t}\leftarrow x_{0|t}-\lambda_t\nabla(R_H+R_G) to the clean estimate at each step (implemented via differentiable Newton completion; λt=λ0(1−αˉt)\lambda_t=\lambda_0(1-\bar{\alpha}_t), λ0=0.01\lambda_0=0.01. The original paper does not specify σt\sigma_t, so we adopt the standard DDIM convention).

PIDSolver = pre-training (denoiser Θ1\Theta_1 + surrogate network Θ2\Theta_2, with the surrogate trained on the bi-objective MSE$+\eta|h|^2$) + fine-tuning (Eq. (28), λh=λg=0.1\lambda_h=\lambda_g=0.1, λc=0.0001\lambda_c=0.0001); inference is deterministic TT-step DDIM, with no clipping/correction post-processing. The final dependent variables z\mathbf{z} are obtained from an independent PyPower AC-PF solve, and the surrogate network exists only in the training stage. The detailed Algorithm can be found in Appendix E.

(B) Regarding the inference-time accounting conventions: To make the conventions checkable, we explicitly state the computation included in the timing of each row:

Method24-bus (s/1k)118-bus (s/1k)300-bus (s/1k)Computation included in the row
M1 (DeepOPF-V)<0.01<0.01<0.01DNN forward + power-balance-reconstruction correction
M2 (PINN)0.050.07--- (did not converge)DNN forward
M3 (DC3 completion stage)<0.01<0.01<0.01DNN forward only (correction diverges within 2–3 steps)
M4 (DDPM)1.491.521.521000-step DDPM denoising
M5 (constrained DDIM)1.523.709.2530-step DDIM denoising + per-step constraint correction (differentiable Newton completion)
PIDSolver1.511.551.541000-step DDIM denoising
PyPower OPF (reference)163.9346.0907.6per-instance AC-OPF solve

(C) Regarding M2's training acceleration: We reproduced the "batched GPU solve" version recommended in the original paper (the completion Newton system is vectorized across the batch, performing one (512,n,n)(512,n,n) linear solve per sweep), and updated the training-time statistics accordingly. To align with the naive "per-sample Newton + implicit-function-theorem gradient" implementation in DC3, we also measured the per-sample serial Newton method: under the same 3000 training epochs, the 118-bus system is estimated to take about 88 hours. We should also state honestly: in our reproduction, M2 did not converge on the 300-bus system (reflecting a scale limitation of the method and the original paper did not test the 300-bus system). The updated offline training-time comparison on a single RTX 5090 GPU is summarized below (as shown in Table 3):

Method24-bus118-bus300-bus
M1 (DeepOPF-V)$\le$5 min$\le$5 min$\le$5 min
M2 (PINN)$\le$1 h$\le$1 h---
M3 (DC3)---
M4 (DDPM)$\le$1 h$\le$1 h$\le$1 h
M5 (constrained DDIM)$\le$1 h$\le$1 h$\le$1 h
PIDSolver$\le$1 h$\le$1 h$\le$1 h

PIDSolver training cost is also comparable with the batched-GPU version of the PINN-based solver M2, whereas the serial per-sample Newton protocol of M2 extrapolates to $\approx$88,h on 118-bus system. M3 is reported as ``-'' because its correction phase diverges within 2--3 epochs on all three systems and its training never completes. Its projected full-budget cost is $\approx$86,h on 118-bus system, confirming that the per-sample solver-plus-inversion channel is unaffordable at diffusion-training scale.

Weakness 2: Regarding the speed–feasibility trade-off, Table 2’s batch GPU times lack a matched comparison with conventional OPF solving under comparable tolerances.

Response. We thank the reviewer for this insightful comment. We have now added a matched comparison with conventional OPF solvers and per-sample feasibility statistics, all measured under the same feasibility conventions, as explained below.

(A) Regarding the comparison with conventional OPF solvers under matched conventions: We report the latency of solving a single AC-OPF instance with Ipopt (within the PowerModels framework, with no solver options modified, i.e., the Ipopt 3.x default convergence tolerances: tol =10−8=10^{-8}, acceptable_tol =10−6=10^{-6}). On the same machine (PIDSolver on an RTX 5090 GPU, Ipopt on its CPU), the inference times are listed item by item below:

SystemIpopt single instance (ms, median)PIDSolver single instance (s, denoising n=1, measured)PIDSolver 1000-sample batch (s, denoising)Ipopt 1000 samples total
2441.5≈0.93\approx0.931.51≈\approx 40
118166.5≈0.97\approx0.971.55≈\approx 170
300487.1≈0.93\approx0.931.54≈\approx 500

We would also like to explain to the reviewer the significance of this comparison in practice: real power dispatch requires repeatedly solving AC-OPF over a vast number of load scenarios (in day-ahead and real-time dispatch the number of scenarios runs into the hundreds of thousands, and the cumulative solve time is enormous). The value of PIDSolver lies precisely in amortizing the optimization into training, delivering dispatch solutions close to AC-OPF quality with second-scale batched inference. Therefore, comparing only single-instance solve times underestimates the practical significance of this batched, amortized solving. We present the numbers of both sides honestly, side by side, for the reviewer to judge this trade-off.

(B) Regarding whether low average violations imply feasible outputs: Low average violations alone do not guarantee per-sample feasibility, so we introduce per-sample feasibility statistics over the 1000 test samples. A sample is counted as feasible if, after an independent PyPower solve, none of the four violation classes exceeds the threshold τ\tau. Samples for which the AC-PF solve fails are counted as infeasible, and the denominator is always the full set of 1000 samples. How is the threshold chosen? In engineering practice, dispatchers often resort to direct current optimal power flow (DC-OPF), the linearized simplification of the AC model, to trade accuracy for speed. Although AC-OPF has an optimal solution, solving each instance from scratch is too slow for real-time dispatch. One therefore has to resort to a DC-OPF surrogate, even though the resulting solutions are inexact. We therefore set τ\tau by taking the violation results of the DC-OPF solution as the reference and shrinking their magnitude by 10×10\times, which gives τ=5×10−3,5×10−3,2×10−2\tau=\\{5\times10^{-3},5\times10^{-3},2\times10^{-2}\\} for the three test cases.

Method24: 5×10−35\times10^{-3}118: 5×10−35\times10^{-3}300: 2×10−22\times10^{-2}
M1 (DeepOPF-V)41.7%7.2%47.2%
M2 (PINN)64.1%0.0%--- (did not converge)
M3 (DC3 completion stage)3.5%9.3%18.0%
M4 (DDPM)74.9%13.9%58.3%
M5 (constrained DDIM)62.7%10.0%50.2%
PIDSolver100.0%100.0%99.4%

Under their respective discriminating thresholds, PIDSolver achieves per-sample feasibility fractions of 100% on the 24/118-bus systems and 99.4% on the 300-bus system, while the strongest baseline only reaches 74.9%/13.9%/61.9% (M4 on 24/118, M3 on 300) — the lower average violations likewise translate into an overwhelming feasibility advantage at the per-sample level. The detailed results and anlysis can be found in Appendix G.6.

(C) Regarding the independent verification procedure: This revised manuscript now states clearly that all reported metrics are derived from independent PyPower AC-PF computations, independent of the surrogate model.

Reviewer_kgD4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ANN-based solvers for problems with implicit algebraic constraints, such as AC-OPF, typically require traditional solvers during training to recover latent variables and estimate gradients through the implicit solution path, resulting in substantial computational overhead. To address this issue, this paper proposes PIDSolver, a sequential physics-informed diffusion solver for optimization problems with complex algebraic constraints. The core of PIDSolver is a sequential architecture in which a noise prediction network is coupled with an algebraic constraint evaluation network. The latter directly predicts the latent variables, enabling the gradients of equality and inequality constraints to propagate end-to-end to the diffusion model and thereby improving training efficiency.

The reviewer agrees with the limitation identified in the paper, and PIDSolver appears to provide an effective solution to this issue. However, the paper suffers from a clear overclaiming problem. In particular, although the title presents PIDSolver as a method for “Optimization with Algebraic Constraints,” the main experiments are restricted to a single IEEE 118-bus AC-OPF system. Therefore, the reviewer is currently unable to properly assess whether the proposed method is broadly effective, and the central claim of the paper is not sufficiently validated.

Critical strengths
  • This paper clearly identifies a key limitation of existing methods for solving AC-OPF, and the proposed PIDSolver directly addresses this issue.
  • On the IEEE 118-bus AC-OPF benchmark, PIDSolver demonstrates substantially better performance than the compared methods.
Critical weaknesses
  • The major concern is overclaiming. Although the title and contribution summary present PIDSolver as a method for “Optimization with Algebraic Constraints,” the main experiments only consider the IEEE 118-bus AC-OPF problem. The paper neither evaluates PIDSolver on AC-OPF systems of different scales nor provides sufficient evidence to demonstrate its applicability to broader classes of algebraically constrained optimization problems. Compared with general constrained-optimization methods such as DC3, the current benchmark coverage is too narrow to support such a broad claim.
  • There are also some confusing aspects regarding the method design. In Equation 10, the operating cost is included as part of the physics-informed objective through an explicit cost-gradient term. However, the final configuration sets λc = 0, and Table D.2 shows that using a positive λc degrades feasibility without further improving the operating cost. Therefore, it remains unclear what practical role the explicit cost-gradient design plays in the final method and whether it should be considered an effective component of PIDSolver.
Questions for authors

Please refer to the weaknesses above.

6 comments

We would like to express our sincere gratitude to the reviewer for the thorough and insightful comments, as well as for the kind recognition of our work, which are very encouraging to us. We have carefully addressed each of the concerns raised, as detailed below. In particular, the comments prompted us to substantially strengthen the empirical evaluation: we extended the main experiments to three AC-OPF systems, re-measured all baselines and PIDSolver under a unified no-clipping protocol, and re-examined the role of the cost term on a much finer weight grid. We sincerely hope that these clarifications and revisions fully resolve the points raised, and we would be most grateful for any further comments.

Weakness 1: Regarding the claims, the broad “Optimization with Algebraic Constraints” scope is unsupported by experiments on only one IEEE 118-bus AC-OPF system.

Response. We sincerely thank the reviewer for carefully pointing out this overclaiming issue. Following the reviewer's valuable suggestion, we have corrected it along two directions, narrowing the claim and providing additional evidence, and we respond point by point below.

(A) On the scope of the title and abstract: The title has been narrowed to "PIDSolver: A Sequential Physics-Informed Diffusion Solver for Alternating Current Optimal Power Flow with Algebraic Constraints," restricting the scope to AC-OPF problems with algebraic constraints

(B) On evaluation on only a single AC-OPF system: We agree with this comment, and we have extended the main evaluation to three AC-OPF systems (24/118/300 buses; six learning methods in total — PIDSolver and five baselines M1–M5; mean±std over multiple training seeds per method. The details results and analysis can be found in Table.1 and Paragraph 1 of Section 4.3. On all systems, PIDSolver achieves the lowest lql_q among the learning methods.

Method24 cost gap (%)24 l_p (×10⁻²)24 l_q (×10⁻²)24 l_u (×10⁻²)24 time (s/1k)118 cost gap (%)118 l_p (×10⁻²)118 l_q (×10⁻²)118 l_u (×10⁻²)118 time (s/1k)300 cost gap (%)300 l_p (×10⁻²)300 l_q (×10⁻²)300 l_u (×10⁻²)300 time (s/1k)
M1 (DeepOPF-V)+3.09±0.030.63±0.160.31±0.060.02±0.00<0.01+0.11±0.020.83±0.040.51±0.020.00±0.00<0.01+0.17±0.031.91±0.121.14±0.080.00±0.00<0.01
M2 (PINN)+3.40±0.380.00±0.000.43±0.030.21±0.040.05+4.41±0.252.75±0.263.95±0.620.01±0.000.07--- (did not converge)------------
M3 (DC3 completion stage)+0.33±0.450.32±0.083.99±2.160.08±0.03<0.01+0.07±0.020.03±0.011.51±0.280.01±0.00<0.01+0.17±0.001.17±0.032.38±0.800.01±0.00<0.01
M4 (DDPM, no physics)-0.02±0.000.03±0.000.36±0.000.01±0.001.49+0.01±0.000.01±0.001.14±0.000.00±0.001.52+0.05±0.000.74±0.012.02±0.030.01±0.001.52
M5 (constrained DDIM correction)-0.00±0.000.02±0.000.47±0.010.01±0.001.52+0.00±0.000.01±0.001.09±0.010.00±0.003.70+0.05±0.000.63±0.012.43±0.080.01±0.009.25
PIDSolver+0.01±0.010.02±0.010.04±0.020.01±0.001.51+0.00±0.000.00±0.000.01±0.000.01±0.001.55+0.51±0.020.05±0.001.05±0.110.00±0.001.54

(C) On evidence for applicability to broader algebraic constraints beyond AC-OPF: Rather than deleting the claim, we support it empirically by transferring PIDSolver to an economic dispatch problem with AGC dynamic constraints. The symbol roles correspond exactly to AC-OPF, but the implicit constraint becomes a differential equation, and the optimization problem is written as (The details can be found in the Paragraph 6 of Section 4.4 and the Appendix A) min⁡x0  C(x0,y)s.t.h(x0,y,z)=0,g(x0,z)≤0,\min_{\mathbf{x}_0}\; C(\mathbf{x}_0,\mathbf{y}) \quad \text{s.t.} \quad h(\mathbf{x}_0,\mathbf{y},\mathbf{z})=0,\quad g(\mathbf{x}_0,\mathbf{z})\le 0, h:s˙=As+B1d+B2u,s(0)=0.h:\quad \dot{\mathbf{s}}=A\mathbf{s}+B_1 d+B_2\mathbf{u},\quad \mathbf{s}(0)=0.

Results over 300 evaluation scenarios are shown below:

Methodlcostl_{cost} (%)cap. viol. (MW)u-lim. (%)ramp-lim. (%)Σr−Rdyn\Sigma r{-}R_{\text{dyn}} (MW)ACE RMS gap (%)
LP reference000.205.230.00
Traditional ED0043.873.41−19.7±12.3-19.7\pm12.3+108.4+108.4
PIDSolver0.22±0.130.22\pm0.130.85±0.130.85\pm0.130.02±0.030.02\pm0.034.04±0.204.04\pm0.20+7.4±1.5+7.4\pm1.5+0.2+0.2
M4−0.00±0.01-0.00\pm0.015.00±0.205.00\pm0.201.01±0.311.01\pm0.314.90±0.014.90\pm0.01+6.0±1.4+6.0\pm1.4+0.1+0.1
M5−0.01±0.01-0.01\pm0.0110.00±0.1010.00\pm0.100.54±0.040.54\pm0.044.58±0.024.58\pm0.02+0.0±1.6+0.0\pm1.6+0.1+0.1

PIDSolver learns a dynamics-aware reserve allocation and it is simultaneously best on dynamic and static feasibility. In summary, the applicability claim is not empty: replacing the algebraic equations with differential equations only requires changing Θ2\Theta_2's completion target to a trajectory solver and the physics residual to the discretized-flow residual.

Weakness 2: Regarding the method design, the cost-gradient term’s role is unclear because λc=0\lambda_c = 0 in the final configuration and positive λc\lambda_c degrades feasibility without improving cost.

Response. We sincerely thank the reviewer for this careful and insightful question. Indeed, this issue had escaped our attention in the earlier submission, and we are very grateful to the reviewer for pointing it out with such professional care. We agree that the previous conclusion was an artifact of our overly coarse search grid, which tested only λc∈0,10−3,10−2\lambda_c\in\\{0,10^{-3},10^{-2}\\}. To answer this question fully, we have conducted additional experiments over a much finer weight grid, and the final conclusion is that λc=10−4\lambda_c=10^{-4} achieves the best trade-off between feasibility and cost, while larger or smaller values degrade feasibility. We explain the role of this component in detail below.

(A) On the problem structure of AC-OPF. In AC-OPF, the active-power output and terminal voltage of generators are decision variables x0=(Pg,Vm)\mathbf{x}0=(P_g,V_m), whereas the reactive-power output of generators is a controlled quantity that belongs to the dependent variables z=(Qg,VPQ,θ)\mathbf{z}=(Q_g,V_{PQ},\theta) implicitly defined by the AC-PF equations, and can only be obtained by solving AC-PF. The operating cost depends only on the generator active power PgP_g, so PgP_g is relatively easy to reproduce accurately by supervised learning—indeed, the plain DDPM baseline (M4) with no physics information already performs well on the cost-gap metric, and the cost gaps of the other compared methods are also very small. Violations are almost entirely concentrated on reactive power. Moreover, the reactive power of a generator is determined indirectly by the terminal voltage (and must be obtained via an AC-PF solve), and the AC-PF is highly nonlinear, so it is difficult to control reactive power precisely by learning the terminal voltage alone. In other words, in the data regime in which PIDSolver operates, the real difficulty that physics-informed fine-tuning must overcome is the fine-tuning of reactive-power feasibility—this is precisely the responsibility of the λh\lambda_h and λg\lambda_g channels.

(B) On why λc=0\lambda_c=0 was chosen. Choosing λc=0\lambda_c=0 stems from the previously coarse hyperparameter grid: The earlier manuscript only tested the settings λc∈0,10−3,10−2\lambda_c\in\\{0,10^{-3},10^{-2}\\} and observed that a positive λc\lambda_c monotonically degrades feasibility without improving cost. The mechanism is an order-of-magnitude mismatch: the cost term is λcC/105\lambda_c C/10^5, where C/105≈1.3C/10^5\approx1.3, whereas under λh=λg=0.1\lambda_h=\lambda_g=0.1 the physical-residual term λh∥h∥2+λg∥Rg∥22\lambda_h\|\mathbf{h}\|^2+\lambda_g\|\boldsymbol{\mathcal{R}}_{\mathbf{g}}\|_2^2 differs substantially in magnitude from the cost term. Consequently, restricting the search to λc∈0,10−3,10−2\lambda_c\in\\{0,10^{-3},10^{-2}\\} caused us to miss the optimal setting and led to an erroneous conclusion.

(C) On the best parameter configuration. To answer this question fully, we retrained on the complete grid λh∈0,0.01,0.1×λc∈0,10−5,10−4,10−3,10−2,10−1\lambda_h\in\\{0,0.01,0.1\\}\times\lambda_c\in\\{0,10^{-5},10^{-4},10^{-3},10^{-2},10^{-1}\\}; the resulting lql_q (mean±std across seeds, ×10−2\times10^{-2}) is reported below:

λh  \  λc\lambda_h \;\backslash\; \lambda_c0010−510^{-5}10−410^{-4}10−310^{-3}10−210^{-2}10−110^{-1}
000.019±0.0020.019\pm0.0020.019±0.0020.019\pm0.0020.019±0.0020.019\pm0.0020.020±0.0030.020\pm0.0030.032±0.0120.032\pm0.0120.372±0.172†0.372\pm0.172^{\dagger}
0.010.010.024±0.0040.024\pm0.0040.017±0.0040.017\pm0.0040.017±0.0040.017\pm0.0040.019±0.0040.019\pm0.0040.041±0.0040.041\pm0.0040.069±0.003†0.069\pm0.003^{\dagger}
0.10.10.015±0.0030.015\pm0.0030.011±0.0030.011\pm0.0030.011±0.003\mathbf{0.011\pm0.003}0.012±0.0030.012\pm0.0030.024±0.0040.024\pm0.0040.074±0.007†0.074\pm0.007^{\dagger}

The grid therefore verifies that {λc,λh,λg}=10−4,0.1,0.1\{\lambda_c,\lambda_h,\lambda_g\}=\\{10^{-4},0.1,0.1\\}—the configuration adopted in the revised paper—is a reasonable trade-off. All of our experiments were rerun under this configuration. More detailed parameter experiments are reported in Appendix D and Appendix G.2.

Reviewer_Kg3jOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes PIDSolver, a physics-informed diffusion framework evaluated on alternating current optimal power flow (AC-OPF). The method couples a denoising network with a learned power-flow evaluation network, allowing constraint penalties to propagate through the predicted dependent variables without repeatedly invoking a numerical power-flow solver during training. The networks are pretrained and then jointly fine-tuned. Experiments on the IEEE 118-bus system report reduced constraint violations while maintaining generation costs close to the reference solver.

Critical strengths
  • The paper addresses a relevant computational challenge: incorporating constraints involving implicitly determined variables into diffusion-model training.
  • The reported reduction in reactive-power constraint violations relative to the diffusion baselines is promising.
  • The sampling-budget analysis, fine-tuning ablation, and sensitivity studies provide useful information about the method’s empirical behavior.
Critical weaknesses
  1. The incremental methodological contribution is insufficiently established. The work currently reads primarily as an application and integration of existing physics-informed diffusion techniques for AC-OPF. Prior work, particularly Bastek et al., Physics-Informed Diffusion Models, ICLR 2025, already incorporates physics residuals into diffusion training, discusses inequality constraints and optimization objectives, and uses accelerated DDIM sampling to estimate clean samples for residual evaluation. The learned power-flow network and its joint training with the denoiser appear to be the most distinctive aspects of this submission. However, the paper does not sufficiently isolate the new methodological insight or establish the added value of this coupling relative to existing physics-informed diffusion approaches. The fine-tuning ablation demonstrates that physics-informed training helps, but does not by itself isolate the contribution of replacing numerical power-flow completion with the learned network. An application-focused contribution could be valuable, but its technical or empirical advance needs to be articulated more clearly.

  2. The scope of the claims exceeds the demonstrated evidence. The title and framing suggest a broadly applicable solver for optimization with algebraic constraints, and the introduction additionally claims applicability to differential equality constraints. However, the substantive evaluation concerns one AC-OPF benchmark, supplemented by low-dimensional illustrative examples. These results provide evidence for the proposed AC-OPF approach, but offer limited support for broader claims about constrained optimization. The contribution list should distinguish the proposed network coupling from adopted components such as physics-informed penalties and accelerated sampling. It should also separate demonstrated capabilities from potential extensions. I would encourage the authors either to narrow the claims to the problem class and setting actually studied, or to provide additional justification and evidence for the broader scope.

  3. The inequality penalty is inconsistent with the problem formulation. Equation (3) defines inequality constraints as g(x0,z)≤0g(x_0,z)\leq 0, whereas Equation (10) penalizes ∥g(x0∗,z∗)∥2\|g(x_0^*,z^*)\|^2. As written, this penalizes strictly feasible negative values and pushes them toward the constraint boundary. Equation (5) already defines this positive-part residual, and Appendix B uses ReLU penalties. This may therefore be a notation inconsistency rather than an implementation error. Nevertheless, the authors should clarify which penalty was used in the AC-OPF experiments and ensure that the loss, gradients, and feasibility statements use a consistent definition.

  4. The complete inference procedure is insufficiently specified. Section 3.3 provides the deterministic sampling update and states that it is used at inference. However, Equation (13) contains only the noise-prediction network, while the abstract describes the second network as performing gradient-based corrections. The paper should provide an explicit inference algorithm clarifying when the power-flow evaluation network is used, whether any constraint-gradient correction or feasibility restoration occurs during sampling, and how dependent variables are obtained for the final solution. It should also clarify whether feasibility is evaluated using the learned surrogate or an independent numerical power-flow calculation, and which operations are included in the reported inference times.

Questions for authors
  1. What specific methodological advance distinguishes PIDSolver from existing physics-informed diffusion methods combined with a differentiable surrogate?
  2. Does the AC-OPF implementation use gg or its positive part in Equation (10)?
  3. Can the authors provide complete inference pseudocode resolving the roles of the two networks and any constraint corrections?
13 comments

We would like to express our sincere gratitude to the reviewer for the thorough and insightful comments, as well as for the kind recognition of our work, which are very encouraging to us. We have carefully addressed each of the concerns raised, as detailed below, and we sincerely hope that our clarifications and revisions fully resolve the points raised. We would be most grateful for any further comments.

Weakness 1: Regarding the methodology, the novelty is not sufficiently isolated from existing physics-informed diffusion work, nor is the coupling’s added value established.

Response. We thank the reviewer for this detailed and critical comment. We fully agree that "the paper does not sufficiently isolate the new methodological insight," and to this end we have added dedicated decomposition and controlled experiments. We respond point by point below.

(A) On the distinction from existing physics-informed diffusion methods (in particular Bastek et al.et\ al., ICLR 2025): The most important difference lies in the forms of the constraints. In the problems addressed by [Bastek et al.et\ al.], the constraints are explicit, i.e., z=f(x,y)z=f(x,y) has a closed-form analytical expression, so computing zz and its partial derivatives ∂z/∂x\partial z/\partial x is cheap. In alternating-current optimal power flow (AC-OPF), by contrast, the dependent variables zz are implicitly defined by the nonlinear alternating-current power-flow (AC-PF) equations h(x,y,z)=0h(x,y,z)=0, and no closed-form ff exists. Consequently, if such methods were transplanted directly to AC-OPF, every training sample would incur two costs: (i) numerically solving the AC-PF equations to obtain zz; and (ii) applying the Jacobian inverse Jh−1J_h^{-1} via the implicit function theorem to obtain end-to-end gradients.

We measured both costs of this DC3 scheme (numerical completion + implicit-function-theorem gradients) on the three systems (Details can be found in the Paragraph 4 of Section 4.3 and the Appendix G.1):

SystemAC-PF solve (ms/sample)Jacobian inversion (ms/sample)κ(Jh)\kappa(J_h)divergence rate under 10% noisetotal cost per 512-sample batch
245.950.031.86×1021.86\times10^{2}0/200≈3.1\approx3.1 s
1187.361.622.91×1032.91\times10^{3}0/200≈4.6\approx4.6 s
30012.967.961.19×1051.19\times10^{5}28/200 (14%)≈10.7\approx10.7 s

The table yields three conclusions. First, the per-sample gradient cost of implicit constraints inflates disproportionately on large systems, and the condition number of JhJ_h deteriorates by orders of magnitude, making the implicit gradients unreliable at scale. Second, under noisy intermediate actions the power-flow solve diverges on a non-negligible fraction of large-system samples, interrupting the training pipeline on those samples. Third, PIDSolver's surrogate processes the same batch with no solver call and no inversion, which is precisely the structural difference that the ``numerical completion vs.\ learned completion'' comparison isolates.

(B) On the added value of the learned AC-PF network and its joint training with the denoiser: PIDSolver eliminates both overheads through its learned AC-PF surrogate, which provides fast, differentiable zz and end-to-end gradients without any Jacobian inversion (0.45–0.49 s for the same 512-sample batch on all three systems). Two pieces of causal evidence establish the added value of the coupling itself: (i) Physics-informed fine-tuning of Eq. (28) reduces lql_q relative to the physics-free DDPM baseline M4 by 8.6×/105×/1.9×8.6\times/105\times/1.9\times on the 24/118/300-bus systems, but the injection route is decisive: the DC3 correction stage (per-sample AC-PF solve + Jacobian-inversion injection) diverges within 2–3 epochs on all three systems and its completion-stage results are already worse than M4, whereas PIDSolver injects physics through the learned surrogate, converges stably, and is the best on all systems (lline≤0.01×10−2l_{\text{line}}\le0.01\times10^{-2} on all rows) (Details can be found in the Paragraph 1 of Section 4.3):

Method24 cost gap (%)24 l_p (×10⁻²)24 l_q (×10⁻²)24 l_u (×10⁻²)24 PF118 cost gap (%)118 l_p (×10⁻²)118 l_q (×10⁻²)118 l_u (×10⁻²)118 PF300 cost gap (%)300 l_p (×10⁻²)300 l_q (×10⁻²)300 l_u (×10⁻²)300 PF
M4-0.02±0.000.03±0.000.36±0.000.01±0.00100%+0.01±0.000.01±0.001.14±0.000.00±0.00100%+0.05±0.000.74±0.012.02±0.030.01±0.00100%
M3+0.33±0.450.32±0.083.99±2.160.08±0.03100%+0.07±0.020.03±0.011.51±0.280.01±0.00100%+0.17±0.001.17±0.032.38±0.800.01±0.00100%
PIDSolver+0.01±0.010.02±0.010.04±0.020.01±0.00100%+0.00±0.000.00±0.000.01±0.000.01±0.00100%+0.51±0.020.05±0.001.05±0.110.00±0.00100%

(ii) On an already-trained PIDSolver with both networks frozen, we compare the two gradient routes sample-by-sample: given the load yy and action xx, the surrogate produces zz and it is held fixed, and the gradient from z→xz \to x is computed once by automatic differentiation through the surrogate and once by implicit-function-theorem inversion (−J−1∂r/∂x-J^{-1}\partial r/\partial x), with the time of both complete routes measured (Details can be found in the Paragraph 4 of Section 4.3 and the Appendix G.1):

Systemcos⁡(gautodiff,gIFT)\cos(g_{\text{autodiff}}, g_{\text{IFT}}): fixed z, medianPIDSolver gradient route time (ms/sample)DC3 gradient route time (ms/sample)speedup
240.997 (0.964)≈0.9\approx0.95.986.6×6.6\times
1180.900 (0.893)≈0.9\approx0.98.9810.0×10.0\times
3000.773 (0.700)≈0.9\approx0.920.9223.2×23.2\times

The two routes agree closely on the 24/118-bus systems, confirming that the surrogate carries the same physical information as the numerical inversion; the drop to 0.773 on the 300-bus system reflects the ill-conditioned Jacobian inversion at that scale. Since PIDSolver's route needs no solve and no inversion, joint training trains the denoiser both efficiently and accurately.

(C) On "whether the fine-tuning ablation can, by itself, isolate the contribution of replacing numerical power-flow completion with the learned network": The reviewer's judgment is entirely correct: a simple fine-tuning on/off ablation only shows that physics-informed fine-tuning helps, and cannot by itself isolate the contribution of replacing numerical completion with the learned network. We therefore isolate this contribution directly at two levels. At the efficiency and stability level, numerical completion requires one AC-PF solve plus one Jacobian inversion per training sample, whereas the learned surrogate needs only a 0.45–0.49 s forward+backward pass per batch and no inversion. At the gradient and performance level, the fixed-zz comparison shows that the surrogate's automatic-differentiation gradients carry the same physical information as the implicit-function-theorem inversion. In summary, the learned network replaces numerical completion with equal physical information, much lower overhead, and greater numerical stability, so joint training trains the denoiser efficiently and accurately.

Weakness 2: Regarding the claims, their broad scope exceeds evidence limited to one AC-OPF benchmark and low-dimensional examples.

Response. We sincerely thank the reviewer for this important and constructive comment. Following the reviewer's valuable suggestion, we have adopted both directions, narrowing the claims and providing additional evidence, and we explain them point by point below.

(A) On distinguishing the proposed coupling from adopted components (physics-informed penalties, accelerated sampling): We agree with the reviewer, and in the revised manuscript's contributions paragraph in the introduction these statements have been removed. But we also wish to sincerely clarify to the reviewer that the contribution of this work is not the physics-informed loss function itself, the contribution is to provide a physics-informed gradient link for such losses: PIDSolver obtains end-to-end gradients directly through the learned differentiable surrogate, requiring no solver calls and no matrix inversion.

(B) On the breadth of the title and framing: We have narrowed the title as the reviewer suggested: "PIDSolver: A Sequential Physics-Informed Diffusion Solver for Alternating Current Optimal Power Flow with Algebraic Constraints", explicitly limiting it to the class of AC-OPF problems with algebraic constraints. The abstract and the contribution list have been rewritten accordingly to retain only empirically supported claims.

(C) On the evidence covering only a single AC-OPF benchmark: We agree with this comment, and we have extended the main evaluation to three AC-OPF systems (24/118/300 buses; six learning methods in total — PIDSolver and five baselines M1–M5; mean±std over multiple training seeds per method. The details results and analysis can be found in Table.1 and Paragraph 1 of Section 4.3. On all systems, PIDSolver achieves the lowest lql_q among the learning methods.

Method24 cost gap (%)24 l_p (×10⁻²)24 l_q (×10⁻²)24 l_u (×10⁻²)24 time (s/1k)118 cost gap (%)118 l_p (×10⁻²)118 l_q (×10⁻²)118 l_u (×10⁻²)118 time (s/1k)300 cost gap (%)300 l_p (×10⁻²)300 l_q (×10⁻²)300 l_u (×10⁻²)300 time (s/1k)
M1 (DeepOPF-V)+3.09±0.030.63±0.160.31±0.060.02±0.00<0.01+0.11±0.020.83±0.040.51±0.020.00±0.00<0.01+0.17±0.031.91±0.121.14±0.080.00±0.00<0.01
M2 (PINN)+3.40±0.380.00±0.000.43±0.030.21±0.040.05+4.41±0.252.75±0.263.95±0.620.01±0.000.07--- (did not converge)------------
M3 (DC3 completion stage)+0.33±0.450.32±0.083.99±2.160.08±0.03<0.01+0.07±0.020.03±0.011.51±0.280.01±0.00<0.01+0.17±0.001.17±0.032.38±0.800.01±0.00<0.01
M4 (DDPM, no physics)-0.02±0.000.03±0.000.36±0.000.01±0.001.49+0.01±0.000.01±0.001.14±0.000.00±0.001.52+0.05±0.000.74±0.012.02±0.030.01±0.001.52
M5 (constrained DDIM correction)-0.00±0.000.02±0.000.47±0.010.01±0.001.52+0.00±0.000.01±0.001.09±0.010.00±0.003.70+0.05±0.000.63±0.012.43±0.080.01±0.009.25
PIDSolver+0.01±0.010.02±0.010.04±0.020.01±0.001.51+0.00±0.000.00±0.000.01±0.000.01±0.001.55+0.51±0.020.05±0.001.05±0.110.00±0.001.54

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.