acceptodds
Under review as a conference paper at ICLR 2027

Across RF Cycles: Poisson-Closed Neural Operators for Capacitively Coupled Plasmas

Abstract

Neural operators (NOs) offer efficient surrogates for time-dependent physical systems, but their use in specialized plasma applications remains limited. We develop a tailored NO for parameterized one-dimensional capacitively coupled plasmas (CCPs), which underpin plasma-assisted processes (e.g., semiconductor etching). CCP dynamics are driven by radio-frequency (RF) forcing, producing strongly oscillatory trajectories that relax over many cycles toward a quasi-periodic state. These multiscale, long-horizon dynamics are difficult for conventional autoregressive NOs because prediction errors accumulate across repeated RF cycles. We address these challenges through a Poisson-closure rollout, Floquet–Poincare-inspired regularization, and random subsequence training for scalable long-trajectory learning. We introduce a parameterized 1D CCP problem as a benchmark for developing and evaluating NOs.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural operator, autoregressive rollouts, scientific machine learning, AI4science, plasma physics

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_bCw7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Paper proposes a neural operator for long-term prediction of capacitively coupled plasma dynamics which combines numerical Poisson solver, an auxiliary cycle-to-cycle predictor, and random subsequence training to reduce prediction errors.

Critical strengths
  1. Using a Poisson solver to reconstruct the electric potential is physically motivated.

  2. The experiments include component comparisons, different trajectory lengths, and temporal extrapolation beyond the training horizon supporting the main claims.

Critical weaknesses
  1. Methodological novelty needs clearer discussion.

  2. Experiments are limited to a 1D geometry with the same initial condition across trajectories. Generalization over varying parameters remains unclear.

  3. Training description has an indexing ambiguity. It uses 10 input states and 118 rollout steps, while the auxiliary target is shifted by 128 steps. Can you clarify it?

  4. Main text and appendix appear to use different mathematical signs for Poisson’s equation and the electron drift term.

  5. Consider comparing against latest baselines such as Transolver, Latent Neural operator and Latent Mamba Operator.

Questions for authors
  1. What does the proposed cycle-level objective add beyond a standard consistency loss between short-step and long-step predictors? Would an auxiliary predictor with a non-RF-aligned horizon provide similar gains?

  2. Could you clarify the exact time indices used for the consistency loss? Is the local rollout extended beyond the stated 118 prediction steps if yes then how?

  3. How much training and inference time does the Poisson closure add? Compare full rollout cost with both the neural baselines and the reference numerical solver.

  4. Can you report all five training seeds, including the excluded runs, and indicate how often each method fails to train reliably?

  5. How does model perform with unseen initial conditions or RF frequencies? Testing several physical parameters would strengthen the generalization claims.

  6. FP regularization performs worse than Poisson closure alone on short trajectories. Could you explain when it becomes beneficial and how sensitive this is to the consistency-loss weight?

  7. I find the physical motivation useful. However, the novelty, training details, and robustness evidence need to be clearly stated.

9 comments
AuthorsOpenReview ↗

Author rebuttal (1/2)

We thank the reviewer for carefully reviewing the paper and providing constructive feedback. The comments have helped us improve the manuscript. While we are still completing several additional numerical experiments, we would like to provide the responses and clarifications available so far.

[W1] Novelty: Thank you for the comment. As the reviewer f2yy acknowledges in the strengths, we would like to reiterate that the objective of this work is not solely to introduce an entirely new standalone learning primitive, but to better address a challenging scientific application through problem-specific methodological developments and their integration. We believe that this is consistent with the broader notion of contribution described in the ICLR reviewer guidelines, which explicitly recognizes work that draws attention to a new application or problem, better addresses an existing problem, and combinations of methodological and application-driven contributions. All together, these adaptations and a proper integration (random-subsequence training for scalable long-trajectory learning, a hybrid neural--numerical rollout for enforcing the Poisson constraint, and auxiliary one-cycle supervision for mitigating long-horizon drift) enable accurate 48-RF-cycle prediction in a regime where the general-purpose neural-operator baselines considered in our experiments perform poorly.

[W2/Q5] Voltage parameter sweep: Thank you for the comment. We have extended the voltage parameter sweep to include 24 RF cycles and consider hpFNOx and hpFNOx variants. The summary of the results are as follows:

hpFNOx: 1.6632, hpFNOx-BC: 0.2830, hpFNOx-PC: 0.2467, and hpFNOx-FP: 0.1680.

The full details are presented in Appendix C.3.

Initial conditions and operating conditions: Thank you for the comment. We agree that generalization across additional initial conditions and operating parameters is an important direction in the general SciML domain. At the same time, the present study is intentionally formulated as a parameterized surrogate-modeling problem in which trajectories evolve from a prescribed initial condition while physical parameters are varied, which is a standard formulation in reduced-basis and parametric model-reduction methods for time-dependent systems [1,2,3]. The primary benchmark spans a broad three-order-of-magnitude range of the reaction-source amplitude, producing substantially different plasma-density regimes. The RF frequency is fixed at (13.56) MHz, a conventional operating frequency widely used in CCP studies and adopted in established CCP simulation benchmarks [4,5].

We also note that variation of the RF driving condition was already considered in the original submission through a voltage-amplitude sweep over (V_0 in [100,300]). That experiment directly changes the time-dependent boundary forcing and consequently the electric field, sheath dynamics, and charged-particle evolution, which we have further extended this experiment to 24 RF cycles in revision. We show that the benefit of FP regularization persists under varying RF-drive amplitudes.

We again agree with the reviewer on the importance of the raised question and we note that generalization beyond this remains beyond the present scope.

[1] Benner, Gugercin, and Willcox, A Survey of Projection-Based Model Reduction Methods ..., SIAM Review (2015)

[2] Hesthaven, Pagliantini, and Rozza, Reduced Basis Methods for Time-Dependent Problems, Acta Numerica (2022)

[3] Grepl and Patera, A Posteriori Error Bounds for Reduced-Basis Approximations of ..., M2AN (2025)

[4] Turner, et al, Simulation benchmarks for low-pressure plasmas: capacitive discharges, Physics of Plasmas, 2013

[5] Chabert and Braithwaite, Physics of Radio-Frequency Plasmas, Cambridge University Press, 2011

[W3/Q2] Time indexing for the consistency loss: Thank you for pointing this out. The reviewer is right. The description on how that consistency loss is computed was not complete in the original submission. We revised the manuscript to clearly state this in Appendix B.2 (p. 25).

In short, when the standard local rollout of length L_sub does not reach one full RF period, we continue the hpFNOx-PC backbone autoregressively for the additional steps needed to reach the phase-aligned one-cycle target. In our setting, this requires one extra rollout step beyond the standard 118-step local rollout.

[W4] sign error: Thank you for the careful reading. This has been fixed; the one in Appendix is correct.

[W5] Additional Baselines: Thank you for the suggestion. We have added a new set of results obtained from LNO (Wang and Wang, NeurIPS, 2024) and LaMO (Tiwari, et al, ICML, 2025) in the Appendices (B.1 and C.2). In summary, the proposed methods remain the best-performing methods, outperforming these additional recent NO baselines.

AuthorsOpenReview ↗

Author rebuttal (2/2)

[Q1] cycle-to-cycle operator: Thank you for the comment. Our motivation for choosing a lag of exactly one RF period is that the CCP trajectories can be viewed as rapidly RF-periodic oscillations modulated by a much slower cycle-to-cycle relaxation. By mapping states separated by (T_RF), the auxiliary operator always compares states at the same forcing phase, thereby largely factoring out the fast intra-cycle oscillation and focusing the learned mapping on the slower evolution of the cycle-wise envelope. Intuitively, this is analogous to sampling a slowly decaying oscillatory signal once per period: the resulting sequence exposes the underlying decay trend much more directly than samples taken at arbitrary phase offsets. This role of the cycle-scale operator is also reflected in the cycle-convergence results (Figure 5), where FP more accurately captures the gradual cycle-to-cycle evolution toward the converged oscillatory state.

So far, we have performed an ablation to distinguish the effect of the learned auxiliary target from direct one-cycle ground-truth supervision (suggested by the reviewer f2yy). On the 24-RF-cycle trajectories, using the ground-truth state at the same one-cycle offset increases the overall error from (0.0333, hpFNOx-FP) to (0.0573, about 42% increase in error). This result suggests that the learned cycle-scale operator provides a useful auxiliary supervisory signal for the local rollout.

We are currently completing additional lag-based ablations to further isolate the role of the one-RF-period timescale relative to more generic longer-timescale supervision. We will incorporate all the ablation results to the revision when the experimental results are ready.

[Q3] Computation: Thank you for the suggestion. We now have added a new section in Appendix C.1.5 reporting the computational timings of the unconstrained version hpFNOx and two variants hpFNOx-PC and hpFNOx-FP. In summary, the variants provide faster inference in practice (7.7~7.8 compared to 8.5 of hpFNOx)---note that the discrete Poisson operator and RF boundary schedule are fixed, the corresponding discrete Green operator (i.e., via a precomputed factorization of the discrete Poisson operator) can be constructed once before the autoregressive rollout. We also report the computational timing of the reference numerical solver (approximately, 120 seconds), supporting the benefits in the multi-query scenarios.

[Q4] Reporting all five runs: Thank you for the comment. We have added a new section in Appendix C.1.4 reporting the results of all five runs, including the originally excluded runs, across RF-cycle counts from 16 to 48. The results show increased variance when the excluded runs are included; however, the overall ranking remains unchanged, with hpFNOx-PC and hpFNOx-FP continuing to provide the best overall performance.

[Q6] FP regularization details Thank you for the question. The FP regularization is intended primarily to suppress cycle-to-cycle error accumulation rather than to improve short-horizon prediction. For short trajectories, the accumulated rollout drift remains relatively small, leaving less room for the additional cycle-scale supervision to provide a measurable benefit over Poisson closure alone. As the rollout horizon increases, however, cycle-to-cycle drift becomes more pronounced, and FP becomes increasingly competitive and beneficial; in our experiments, this behavior emerges from approximately 24 RF cycles onward, with clearer improvements at longer horizons.

Regarding the loss-weight, we want to note that the same consistency-loss weight, 0.01, is used across all RF-cycle lengths and the voltage-sweep experiments without setting-specific tuning, and consistently provides stable improvements in the longer-horizon regime. We are currently completing additional experiments on the sensitivity to the consistency-loss weight and will report those results in the revised manuscript.

I thank the authors for their detailed response, clarifications, and revisions to the manuscript. As I am less familiar with plasma physics, I have one remaining question: does the plasma dataset exhibit nonstationary behavior across the different conditions studied? If so, could the authors explain how the Fourier Neural Operator captures this behavior, particularly any localized high-frequency features or sharp transitions, given the potential limitations of retaining a finite number of Fourier modes?

Reviewer_f2yyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the long-horizon prediction of radio-frequency (RF)-driven capacitively coupled plasmas (CCPs) using neural operators. The system involves coupled electron and ion dynamics, a Poisson constraint, and multiscale behavior characterized by rapid oscillations within each RF cycle and slower cycle-to-cycle relaxation.

The authors propose three improvements to the existing hpFNOx architecture. First, they introduce a Poisson-closure rollout, where the neural operator predicts electron density, ion density, and ion velocity, while the electric potential is reconstructed by solving the Poisson equation. Second, they introduce Floquet–Poincaré-inspired (FP) regularization, which trains an auxiliary cycle-to-cycle operator and minimizes its discrepancy with the local autoregressive rollout. Finally, they use random temporal subsequence sampling to reduce the computational cost of training on long trajectories. The authors also extend the previous simplified CCP benchmark to a four-field fluid model and report improved long-horizon prediction accuracy.

Critical strengths
  1. The problem addressed in this paper is of practical significance. CCP systems involve a Poisson constraint, RF-periodic forcing, and dynamics at different time scales, making long-horizon autoregressive prediction challenging. Compared with the previous simplified CCP benchmark, this work considers more complete electron and ion dynamics and extends prediction to thousands of time steps. I consider this a meaningful application problem.

  2. The method makes reasonable use of the known physical structure. The Poisson-closure rollout enforces the elliptic constraint through a numerical Poisson solve, rather than requiring the network to predict the electric potential directly. Similarly, FP regularization uses the fact that states separated by one RF period correspond to the same phase of the external forcing, providing phase-aligned one-cycle supervision for the local operator. Although these choices are relatively simple, they address specific difficulties in long-horizon prediction without changing the hpFNOx backbone.

  3. The experiments show substantial improvements in long-horizon prediction. The proposed hpFNOx-FP achieves lower prediction errors than the original hpFNOx and the other variants in the main experiments. Figure 5 shows that FP regularization better captures the cycle-to-cycle convergence behavior, while Figure 6 shows that its advantage becomes more pronounced for longer trajectories. The temporal extrapolation results in the appendix provide further supporting evidence. Overall, the results demonstrate the effectiveness of the proposed method for the CCP problem considered here.

Critical weaknesses
  1. The methodological novelty is relatively limited. The paper builds directly on the hpFNOx architecture proposed by Jing et al. (2026). The Poisson closure, which reconstructs the electric potential using a numerical solver, is closely related to existing neural-numerical hybrid solvers and constrained dynamics methods. Random temporal subsequence sampling is also an established training strategy. For the main FP regularization component, Deep-OSG (Chen and Wu, 2023) has studied consistency between evolution operators at different time scales, while Bramburger and Kutz (2020) used data-driven Poincaré maps to study multiscale dynamics. These works are not equivalent to the proposed method, particularly since this paper considers a periodically forced system. Nevertheless, the main ideas behind the proposed components have close precedents in the literature. The new contribution lies mainly in constructing an auxiliary one-cycle operator based on the known RF period and combining these techniques for CCP dynamics. While this is a reasonable problem-specific design, I find the independent methodological contribution relatively weak. In particular, it remains unclear what the auxiliary cycle-to-cycle operator and FP consistency loss add beyond more general supervision across different time scales. More importantly, the paper does not provide a direct comparison with a simpler baseline that applies one-period endpoint supervision to the local autoregressive rollout without introducing an additional learned operator. Consequently, it is difficult to determine whether the observed improvement primarily arises from the physically motivated phase-aligned supervision or from the specific auxiliary-operator consistency mechanism.

  2. The experiments are restricted to a relatively narrow physical setting. Although the paper considers four coupled physical fields and long autoregressive rollouts, the main experiments are conducted on a one-dimensional CCP system with identical initial conditions and fixed RF voltage amplitude and frequency. The main physical parameter varied is the reaction source coefficient R0R_0. The appendix includes temporal extrapolation and a voltage parameter sweep, but the latter only considers four RF cycles. Therefore, the results mainly establish the effectiveness of the method for the particular CCP setting considered here. Moreover, as discussed in the paper's limitations, the current Poisson closure is tied to a particular spatial discretization and geometry. It remains unclear whether the reported long-horizon improvements would hold under more complex CCP operating conditions or in other periodically forced systems.

Questions for authors
  1. Could the authors clarify the methodological contribution of FP regularization relative to existing multiscale operator-consistency methods, such as Deep-OSG, and data-driven Poincaré map approaches? In particular, since ground-truth states one RF period ahead are already available during training, could the authors compare their auxiliary-operator consistency loss with directly applying a one-period endpoint supervision loss to the local autoregressive rollout, without training an additional auxiliary operator? Such an ablation would help isolate the benefit of the proposed FP formulation beyond phase-aligned supervision alone.

  2. Could the authors further discuss the performance of the method under different CCP operating conditions? In particular, the voltage parameter sweep in the appendix only considers short trajectories. Does FP regularization still provide similar improvements in long-horizon prediction when the RF driving parameters or initial conditions are varied?

8 comments
AuthorsOpenReview ↗

Author rebuttal (1/2)

We thank the reviewer for carefully reviewing the paper, pointing out the strengths of the papers, and providing constructive feedback. The comments have helped us improve the manuscript. While we are still completing several additional numerical experiments, we would like to provide the responses and clarifications available so far.

[W1/Q1] Missing related work and connections: Thank you for pointing us to this relevant work. They are very helpful.

In Bramburger and Kutz (2020), they proposed to directly learn the Poincare map using SINDy and use the learned map itself as a reduced dynamical model. Their study considers canonical systems for which Poincaré sections can be explicitly defined and validated (e.g., periodically forced RC circuit admitting an analytical Poincare map). In contrast, our one-RF-cycle operator is introduced only as an auxiliary training-time model. It provides phase-aligned cycle-scale supervision to the fine-scale autoregressive neural operator and is not used for inference.

Regarding Deep-OSG: we agree that autoregressive NOs can be viewed as fixed-step flow-map approximators; however, unlike Deep-OSG, our method does not learn a time-parameterized evolution operators (so called, data-driven flow-maps) or enforce semigroup consistency. We have added a discussion clarifying that our approach instead uses two prescribed evolution scales, with the one-RF-period map serving only as an auxiliary training signal.

In summary, we have added two new paragraphs describing these work and the connection/distinction from our approach in the Appendix B.2 (p 26). Given the main-text space constraint, we plan to incorporate a concise version into the main text in the camera-ready version if accepted.

Novelty: Thank you for the comment. As the reviewer also acknowledges in the strengths, we would like to reiterate that the objective of this work is not solely to introduce an entirely new standalone learning primitive, but to better address a challenging scientific application through problem-specific methodological developments and their integration. We believe that this is consistent with the broader notion of contribution described in the ICLR reviewer guidelines, which explicitly recognizes work that draws attention to a new application or problem, better addresses an existing problem, and combinations of methodological and application-driven contributions. All together, these adaptations and a proper integration (random-subsequence training for scalable long-trajectory learning, a hybrid neural-numerical rollout for enforcing the Poisson constraint, and auxiliary one-cycle supervision for mitigating long-horizon drift) enable accurate 48-RF-cycle prediction in a regime where the general-purpose neural-operator baselines considered in our experiments perform poorly.

Ablation: Our motivation for choosing a lag of exactly one RF period is that the CCP trajectories can be viewed as rapidly RF-periodic oscillations modulated by a much slower cycle-to-cycle relaxation. By mapping states separated by (T_RF), the auxiliary operator always compares states at the same forcing phase, thereby largely factoring out the fast intra-cycle oscillation and focusing the learned mapping on the slower evolution of the cycle-wise envelope. Intuitively, this is analogous to sampling a slowly decaying oscillatory signal once per period: the resulting sequence exposes the underlying decay trend much more directly than samples taken at arbitrary phase offsets. This role of the cycle-scale operator is also reflected in the cycle-convergence results (Figure 5), where FP more accurately captures the gradual cycle-to-cycle evolution toward the converged oscillatory state.

We have performed an ablation to distinguish the effect of the learned auxiliary target from direct one-cycle ground-truth supervision. On the 24-RF-cycle trajectories, using the ground-truth state at the same one-cycle offset increases the overall error from (0.0333, hpFNOx-FP) to (0.0573, about 42% increase in error). This result suggests that the learned cycle-scale operator provides a useful auxiliary supervisory signal for the local rollout.

We are currently completing additional lag-based ablations to further isolate the role of the one-RF-period timescale relative to more generic longer-timescale supervision. We will incorporate all the ablation results to the revision when the experimental results are ready.

AuthorsOpenReview ↗

Author rebuttal (2/2)

[W2/Q2] Voltage parameter sweep: Thank you for the comment. We have extended the voltage parameter sweep to include 24 RF cycles and consider hpFNOx and hpFNOx variants. The summary of the results is as follows:

hpFNOx: 1.6632, hpFNOx-BC: 0.2830, hpFNOx-PC: 0.2467, and hpFNOx-FP: 0.1680.

The full details are presented in Appendix C.3.

Poisson closure tied to spatial discretization: Thank you for the constructive comment. We further extend our limitation section (Appendix D) by adding potential future directions with more detailed ideas. For handling general geometry, we suggest that using mesh- and resolution-indepedent Poisson solvers (such RBF-based or point-cloud-based ones) could mitigate this limitation. Moreover, for the present 1D problem, we additionally perform a controlled experiments with a simplified version of the CCP problem considered in prior work (Jing et al, ICLR 2026)---we choose this to present results within the limited time period of rebuttal. Specifically, we use numerical interpolation and projection inside the Poisson solver so that models trained at a single resolution (Nx=256, spatial domain) can be deployed to different resolutions (Nx=512, 1024) while maintaining reasonable accuracy. For example, with the proposed hpFNOx-PC trained at Nx=256 achieves errors of 2.48E-4, 4.72E-4 and 6.41E-4 at Nx={256,512,1024}. The level of the error degradation is comparable to those of general purpose neural operators. This result hints that the proposed methods can be further extended to handle grid-transfer operations.

Initial conditions and operating conditions: Thank you for the comment. We agree that generalization across additional initial conditions and operating parameters is an important direction in the general SciML domain. At the same time, the present study is intentionally formulated as a parameterized surrogate-modeling problem in which trajectories evolve from a prescribed initial condition while physical parameters are varied, which is a standard formulation in reduced-basis and parametric model-reduction methods for time-dependent systems [1,2,3]. The primary benchmark spans a broad three-order-of-magnitude range of the reaction-source amplitude, producing substantially different plasma-density regimes. The RF frequency is fixed at (13.56) MHz, a conventional operating frequency widely used in CCP studies and adopted in established CCP simulation benchmarks [4,5].

We also note that variation of the RF driving condition was already considered in the original submission through a voltage-amplitude sweep over (V_0 in [100,300]). That experiment directly changes the time-dependent boundary forcing and consequently the electric field, sheath dynamics, and charged-particle evolution, which we have further extended this experiment to 24 RF cycles in revision. We show that the benefit of FP regularization persists under varying RF-drive amplitudes.

We again agree with the reviewer on the importance of the raised question and we note that generalization beyond this remains beyond the present scope.

[1] Benner, Gugercin, and Willcox, A Survey of Projection-Based Model Reduction Methods ..., SIAM Review (2015)

[2] Hesthaven, Pagliantini, and Rozza, Reduced Basis Methods for Time-Dependent Problems, Acta Numerica (2022)

[3] Grepl and Patera, A Posteriori Error Bounds for Reduced-Basis Approximations of ..., M2AN (2025)

[4] Turner, et al, Simulation benchmarks for low-pressure plasmas: capacitive discharges, Physics of Plasmas, 2013

[5] Chabert and Braithwaite, Physics of Radio-Frequency Plasmas, Cambridge University Press, 2011

AuthorsOpenReview ↗

Additional Ablation Results

Dear Review,

We thank again for your careful review and constructive feedback. We believe those are very helpful in improving our manuscript. Here we present the additional results from further ablation studies on hpFNOx-FP.

Varying consistency loss weight: We have examined the sensitivity to the consistency-loss weight. Across the tested values λFP=10−3,10−2,10−1,100\lambda_\mathrm{FP}=10^{-3},10^{-2},10^{-1},10^{0}, we did not observe severe sensitivity: the changes in prediction error were modest (1.4~1.5% changes in most cases) and did not alter the overall ranking among the compared methods.

Varying horizon lags: We have further varied the auxiliary horizon lags, P=112,128,144,160P=112,128,144,160. The resulting prediction errors remained broadly comparable. However, the cycle-convergence behavior shows a much clearer difference. The RF-aligned choice, P=128P=128, closely followed the ground-truth convergence trend, whereas the non-RF-aligned choices deviated at later cycles. This supports the importance of aligning the cycle-level supervision with the RF period, which allows the surrogate models to capture the important physical property of the considered problem (i.e., initial transient state to cycle-converged states), rather than using an arbitrary long-horizon target.

All the detailed results have been added to Appendix C.4 of the revised manuscript.

Reviewer_SPewOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a neural operator surrogate for one-dimensional capacitively coupled plasma (CCP) fluid dynamics under periodic radio-frequency (RF) voltage forcing. Building on the coupled, hypernetwork-conditioned FNO architecture (hpFNOx) of Jing et al. (2026), the authors combine numerical Poisson closure, an auxiliary one-cycle prediction model that regularizes local autoregressive rollouts, and random temporal subsequence training.

Experiments sweeping the reaction-source amplitude show that hpFNOx-FP reduces the sum of four field-wise nRMSEs over 48 RF cycles from 0.28836 (hpFNOx) and 0.18419 (hpFNOx-PC) to 0.05828. Additional experiments examine cycle-to-cycle relaxation, extrapolation to 64 cycles, shorter rollouts, and a separate voltage-amplitude sweep.

Critical strengths
  1. I think the method makes effective use of the problem’s structure. Predicting the transport variables while reconstructing the electric potential through a numerical Poisson solve enforces the electrostatic constraint and prescribed voltage boundary conditions.

  2. The long-horizon improvements appear substantial. At 48 cycles, hpFNOx-FP improves the aggregate metric by approximately 68% over Poisson closure alone and lowers errors in all four fields.

Critical weaknesses
  1. The auxiliary cycle-scale supervision is a central contribution, but its benefit is not sufficiently isolated. The PC–FP comparison does not establish whether the improvement requires a learned auxiliary target or a prediction lag equal to one RF period, rather than longer-timescale supervision more generally.
  2. The evidence for robustness and practical efficiency is incomplete. Appendix B.4 reports that the worst validation run out of five is excluded for every method. This rule is disclosed and does not use test performance, but excluding the worst-performing run limits assessment of initialization robustness.
Questions for authors
  1. Does the improvement specifically depend on predicting one complete RF period ahead?

  2. Could you report individual test errors for all five seeds of the four hpFNOx variants, including the excluded runs? Do the performance ranking and claimed improvement in initialization robustness remain?

  3. What is the wall-clock time for a complete 48-cycle inference rollout of hpFNOx-PC and hpFNOx-FP, including Poisson solves, compared with the numerical reference solver and unconstrained hpFNOx?

4 comments
AuthorsOpenReview ↗

Author rebuttal

We thank the reviewer for carefully reviewing our paper and for the overall positive assessment. The comments have helped us improve the manuscript. While we are still completing several additional numerical experiments, we would like to provide the responses and clarifications available so far.

[W1/Q1] Detailed study on FP-regularization: Thank you for the comment. Our motivation for choosing a lag of exactly one RF period is that the CCP trajectories can be viewed as rapidly RF-periodic oscillations modulated by a much slower cycle-to-cycle relaxation. By mapping states separated by (T_RF), the auxiliary operator always compares states at the same forcing phase, thereby largely factoring out the fast intra-cycle oscillation and focusing the learned mapping on the slower evolution of the cycle-wise envelope. Intuitively, this is analogous to sampling a slowly decaying oscillatory signal once per period: the resulting sequence exposes the underlying decay trend much more directly than samples taken at arbitrary phase offsets. This role of the cycle-scale operator is also reflected in the cycle-convergence results (Figure 5), where FP more accurately captures the gradual cycle-to-cycle evolution toward the converged oscillatory state.

So far, we have performed an ablation to distinguish the effect of the learned auxiliary target from direct one-cycle ground-truth supervision (suggested by the reviewer f2yy). On the 24-RF-cycle trajectories, using the ground-truth state at the same one-cycle offset increases the overall error from (0.0333, hpFNOx-FP) to (0.0573, about 42% increase in error). This result suggests that the learned cycle-scale operator provides a useful auxiliary supervisory signal for the local rollout.

We are currently completing additional lag-based ablations to further isolate the role of the one-RF-period timescale relative to more generic longer-timescale supervision. We will incorporate all the ablation results to the revision when the experimental results are ready.

[W2/Q2] Reporting all five runs: Thank you for the comment. We have added a new section in Appendix C.1.4 reporting the results of all five runs, including the originally excluded runs, across RF-cycle counts from 16 to 48. The results show increased variance when the excluded runs are included; however, the overall ranking remains unchanged, with hpFNOx-PC and hpFNOx-FP continuing to provide the best overall performance.

[Q3] Computation: Thank you for the suggestion. We now have added a new section in Appendix C.1.5 reporting the computational timings of the unconstrained version hpFNOx and two variants hpFNOx-PC and hpFNOx-FP. In summary, the variants provide faster inference in practice (7.7~7.8 compared to 8.5 of hpFNOx)---note that the discrete Poisson operator and RF boundary schedule are fixed, the corresponding discrete Green operator (i.e., via a precomputed factorization of the discrete Poisson operator) can be constructed once before the autoregressive rollout. We also report the computational timing of the reference numerical solver (approximately, 120 seconds), supporting the benefits in the multi-query scenarios.

AuthorsOpenReview ↗

Additional Ablation Results

Dear Review,

We thank again for your careful review and constructive feedback. We believe those are very helpful in improving our manuscript. Here we present the additional results from further ablation studies on hpFNOx-FP.

Varying consistency loss weight: We have examined the sensitivity to the consistency-loss weight. Across the tested values λFP=10−3,10−2,10−1,100\lambda_\mathrm{FP}=10^{-3},10^{-2},10^{-1},10^{0}, we did not observe severe sensitivity: the changes in prediction error were modest (1.4~1.5% changes in most cases) and did not alter the overall ranking among the compared methods.

Varying horizon lags: We have further varied the auxiliary horizon lags, P=112,128,144,160P=112,128,144,160. The resulting prediction errors remained broadly comparable. However, the cycle-convergence behavior shows a much clearer difference. The RF-aligned choice, P=128P=128, closely followed the ground-truth convergence trend, whereas the non-RF-aligned choices deviated at later cycles. This supports the importance of aligning the cycle-level supervision with the RF period, which allows the surrogate models to capture the important physical property of the considered problem (i.e., initial transient state to cycle-converged states), rather than using an arbitrary long-horizon target.

All the detailed results have been added to Appendix C.4 of the revised manuscript.

Thank you for the detailed response and the additional experiments. I think these new ablations strengthen the paper. I will keep my positive score.

Thank you! We appreciate it!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.