acceptodds
Under review as a conference paper at ICLR 2027

NEURAL FIELD-BASED CHARACTERISTIC RECONSTRUCTION FOR VLASOV–POISSON DYNAMICS

Abstract

The particle distribution in Vlasov–Poisson dynamics develops increasingly fine structures over time, making the full distribution difficult for a neural network to represent directly, whereas the electric field driving particle trajectories depends only on position and time and usually remains comparatively smooth. Motivated by this difference, we introduce Neural Field-Based Characteristic Reconstruction (NFCR), in which the neural network represents only the electric field rather than directly learning the full distribution. A time network outputs a finite set of spatial Fourier coefficients over the training interval, and differentiable backward characteristic integration reconstructs the evolving distribution from the prescribed initial condition. The density and current obtained from this distribution are related to the network output through the Poisson and Amp` ere equations, keeping the electric field consistent with the reconstructed distribution; conditions on the initial field and its time derivative constrain the dynamics near the beginning of the interval. The neural network therefore needs to learn only a finite number of electric-field coefficients, while fine distribution scales are generated naturally by transport, without an evolving phase-space grid or reference labels. Backward trajectories are integrated with a symmetric drift–acceleration–drift scheme. This scheme locally preserves phase-space area during backward tracing and ensures that a distribution reconstructed from a nonnegative initial condition remains nonnegative. Experiments on Gaussian transport and Landau damping show that NFCR reconstructs the distribution with lower errors than neural methods that directly learn the distribution or the full particle-trajectory map, while accurately capturing the electric-field evolution and maintaining small mass drift.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Physics-informed machine learning, Vlasov–Poisson equations, kinetic equations, characteristic methods, phase-space reconstruction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_P4H1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Neural Field-Based Characteristic Reconstruction (NFCR) for Vlasov–Poisson dynamics. Instead of directly approximating the full distribution, NFCR learns a relatively smooth, low-dimensional representation of the electric field and reconstructs the full distribution through differentiable backward characteristic integration. Experiments on Gaussian transport and Landau damping demonstrate the effectiveness of NFCR.

The paper is motivated by well-founded physical considerations. The core idea, learning a relatively simple dynamical driving quantity and reconstructing complex states through known physical evolution mechanisms, has potential value for improving PINNs for solving complex dynamical systems. However, the experiments are limited to only two cases within Vlasov–Poisson dynamics, which are insufficient to establish the general applicability of NFCR across more diverse Vlasov–Poisson configurations, let alone the transferability of its core idea to other PDE families. Given this limitation, the reviewer believes that the current experimental evidence is insufficient to substantiate the contribution of this paper as a more general machine learning method.

Critical strengths
  1. The paper effectively exploits the difference in representational complexity between the phase-space distribution and the electric field in Vlasov–Poisson dynamics. By shifting the learning target from the complex distribution to a relatively low-dimensional driving field and reconstructing the full distribution through characteristic integration, NFCR leverages the underlying physical structure. This idea is potentially valuable for exploring more effective neural representations of complex dynamical systems.
  2. Experiments on Gaussian transport and Landau damping demonstrate that NFCR significantly outperforms the compared methods
Critical weaknesses
  1. NFCR is evaluated only on Gaussian transport and Landau damping, without considering more complex initial conditions, strongly nonlinear dynamics, or longer-term evolution. Therefore, the current experiments are insufficient to determine whether the advantages of NFCR can consistently extend to a broader range of Vlasov–Poisson problems.
  2. The proposed method is highly dependent on the specific physical structure of the Vlasov–Poisson system. Although learning a low-complexity driving field and reconstructing complex states through physical dynamics is a potentially generalizable idea, the paper provides neither experimental evidence nor a more general theoretical framework to demonstrate its transferability to other PDE systems. Consequently, the current work appears closer to a specialized neural numerical solver designed for a particular physical system, while its broader methodological contribution to machine learning remains insufficiently established.
Questions for authors
  1. Can the authors further validate the applicability of NFCR through more challenging Vlasov–Poisson experiments? Such experiments would help determine whether the performance advantages of NFCR can be maintained across more diverse and complex Vlasov–Poisson configurations, rather than being limited to the two cases currently considered.
  2. Can the core design principle of NFCR be extended to PDE systems beyond Vlasov–Poisson dynamics? This would help establish that the contribution is not merely a problem-specific design tailored to the physical structure of Vlasov–Poisson dynamics, but a machine learning methodology with broader applicability.
  3. Figure 1 requires further improvement, particularly regarding misaligned arrows, inconsistent fonts, and layout issues.

If the authors can adequately address Question 1, the reviewer would consider increasing the score.

3 comments

Thank you for noting that the experimental coverage, scope of the design principle, and figure presentation need further work. We added a coupled 2D2V experiment and respond to each question below.

Q1: Can the applicability of NFCR be validated on more challenging VP experiments?

We study the initial condition

f₀(x,v)=M₂(v)[1+0.2cos(x₁)+0.2cos(x₁+x₂)]

on the periodic spatial domain [-π,π)² and velocity domain [-6,6)², up to T=4. The two initial perturbations interact to produce genuine two-dimensional coupling. NFCR learns the time-dependent field coefficients and reconstructs the distribution by tracing characteristics back to the initial state; density and current from the reconstructed distribution constrain the field during training. Each of three seeds uses 2000 Adam updates followed by L-BFGS refinement, without reference-solution labels.

The table reports errors averaged over three seeds at nine output times and the model size:

MetricResult
Full-distribution error0.053%
Perturbation-distribution error0.234%
Model fileabout 45 KB

The next table compares perturbation-distribution errors on two velocity-space slices at T=4, together with solution storage:

MethodError on x₁-v₁Error on x₂-v₂Solution storage
SL, 16²×32², h=0.057.28%2.54%8.31 MB
SL, 32²×96², h=0.0250.024%0.012%300.96 MB
NFCR0.17%–0.21%0.14%–0.16%about 45 KB

One NFCR model of about 45 KB represents the full interval. It requires substantially less storage than retaining nine full distribution snapshots, and adding query positions or times later does not increase model size. This experiment demonstrates reconstruction of a four-dimensional distribution with coupled two-dimensional structure.

Q2: Can the core design principle be extended to PDEs beyond VP?

The transferable idea is to learn a simpler driving quantity and reconstruct a complex state from it through known physical processes. For another equation, three conditions are needed: the target state can be computed or differentiably reconstructed from the driver; the initial or boundary state can be evaluated directly or represented compactly; and the physical relation between the driver and reconstructed state can provide a training signal. The new 2D2V experiment supports feasibility for coupled VP dynamics; transfer to other PDEs requires separate validation. We will state these conditions and distinguish the general idea from VP-specific implementations in this work, such as Fourier representation and characteristic backtracing.

Q3: Improving Figure 1

Thank you for pointing out the readability issues in Figure 1. We will standardize fonts, sizes, arrows, and module alignment; clearly show “field network → electric field → characteristic backtracing → distribution reconstruction → density/current → physics-based feedback”; distinguish the reconstruction and training-feedback paths; explain the main symbols in the caption; and check readability at the paper's column width.

Thank you again for helping us strengthen the coupled VP experiment and clarify the design principle.

The reviewer thanks the authors for their response. However, the additional experiments do not sufficiently demonstrate the general applicability of NFCR to VP systems. The response is hard to read, as it not only contains a large amount of detailed description that could be omitted, but also makes extensive use of ";" to connect complex sentences. In addition, although the authors stated that they would revise the paper, no updated version has been uploaded. Upon rereading the paper, the reviewer also found additional presentation issues. For example, the reviewer is unsure why the descriptions before equations do not end consistently: some are followed by a blank space, while others use a colon.

As the reviewer lacks sufficient expertise in the relevant field, the reviewer cannot make a definitive judgment on the substantive contribution of this work. However, the presentation and the deficiencies in the details have significantly increased doubts about the credibility of this work. The reviewer would be willing to give a borderline rating at most.

Thank you for rereading our paper and explaining your concerns. We apologize that our previous response was difficult to read. We briefly address the main points below.

Revised submission and presentation. The revised paper and anonymous code supplement have now been uploaded. We take your comments on presentation seriously. We have revised Figure 1 to improve readability. We have also revised the text introducing equations so that punctuation is consistent with sentence structure.

Core idea and practical value. In VP systems, the distribution develops fine structures that are difficult for a neural network to represent directly. NFCR instead learns the history of the electric field that drives particle motion. It then traces particle trajectories back to the prescribed initial condition to reconstruct the distribution at the requested times and locations. The network therefore represents the field history, while physical transport generates the fine details of the distribution. Training uses physical constraints without reference-solution labels.

For repeated queries at different times and locations within the trained interval, NFCR retains a single model without storing full distribution snapshots in advance. Additional query times or locations do not increase model storage. This representation also supports rapid output after training. In the 1D1V Landau experiment, NFCR generates the full terminal distribution and electric field in 0.0611 seconds. This gives output-stage speedups of approximately 92.0× and 35.7× over NuFI and SL, respectively, when those methods solve from the initial condition.

Evidence from higher-dimensional experiments. The additional 2D2V experiment includes interactions between perturbations in different spatial directions, rather than simply embedding a one-dimensional solution in a higher-dimensional domain. Across three independent seeds, the mean errors in the full distribution and its perturbation are 0.053% and 0.234%, respectively. Figure 5 in the revised paper shows distribution slices in both directions. NFCR stores the entire interval in an approximately 45 KB model. By comparison, nine full distribution snapshots from the tested SL configurations require 8.31–300.96 MB after compression.

In the 3D3V experiment, NFCR reconstructs the six-dimensional distribution using only 2,042 parameters. Its perturbation error is 0.486%, close to NuFI’s 0.463% and below the 7.09% of the tested SL configuration. Terminal output after training also shows advantages in speed and peak GPU memory. The full comparison is included in the revised paper.

These experiments extend the evidence from one-dimensional tests to coupled two-dimensional dynamics and six-dimensional distribution reconstruction. They help clarify NFCR’s reconstruction capability for prescribed initial conditions over finite intervals, together with the practical value of compact storage and rapid on-demand output.

Thank you again for the opportunity to clarify the experimental evidence and intended contribution.

Reviewer_Q4JYOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Neural Field-Based Characteristic Reconstruction (NFCR) for the Vlasov–Poisson system. Instead of directly approximating the phase-space distribution or its inverse flow map, a neural network takes time as input and predicts coefficients of a truncated spatial Fourier representation of the electric field. The distribution is reconstructed by integrating characteristics backward and evaluating the prescribed initial distribution. Training enforces consistency between the predicted field and reconstructed density and current through Poisson and Ampère residuals, together with an initial-derivative constraint and coefficient regularization. Experiments on Gaussian transport and Landau damping report lower errors than the PINN and SL-PINN.

Critical strengths
  1. Clear motivation and an intuitive design. The paper addresses the difficulty of directly learning increasingly fine structures in the distribution by focusing on the comparatively smooth electric field. This physically motivated approach offers an intuitive way to simplify the learning task while reconstructing detailed dynamics through characteristic integration.
  2. Strong empirical accuracy on the evaluated benchmarks. Tables 1 and 2 show that NFCR consistently achieves the lowest reported errors among the evaluated methods.
Critical weaknesses
  1. The individual benefits of the three claimed contributions are not established through controlled ablations. Appendix E only studies Fourier-mode sensitivity, leaving the effects of the field representation, characteristic integrator, and consistency constraints insufficiently isolated. In particular, comparisons with alternative integrators, Poisson-only training, and removal of the initial-derivative constraint are missing. Although the structural properties are mathematically supported, targeted ablations under matched settings are needed to identify which components drive the reported improvements.
  2. Several implementation details needed for reproducibility are insufficiently specified. The paper does not clearly state the endpoints and sampling allocation of the progressive training windows, the velocity quadrature rule and associated nodes and weights, or the value of (M_s) used in the coefficient-smoothing regularizer. It is also unclear how backward trajectories leaving the truncated velocity interval are handled when evaluating the initial distribution. These choices affect the reconstructed moments, training objective, and numerical accuracy, and should be explicitly documented.
  3. Missing comparison with the closest related method. NuFI is acknowledged as a central antecedent in the Introduction but is absent from the experimental comparisons. Since both methods reconstruct distributions through field-driven characteristics, a direct comparison is needed to establish the added value of the neural field representation in terms of accuracy, computational cost, and memory usage.
Questions for authors
  1. Could you provide controlled ablations of the characteristic integrator, Ampère loss, and initial-derivative constraint to clarify which components drive the improvements under matched training settings?
  2. Could you clarify the progressive training windows and sampling strategy, velocity quadrature settings, the value of (M_s), and the treatment of backward trajectories leaving the truncated velocity interval?
  3. Could you compare NFCR directly with NuFI under comparable numerical settings to establish the added value of the neural field representation in accuracy, computational cost, and memory usage?
2 comments

Thank you for recommending that we examine each component separately, provide reproducibility details, and compare directly with NuFI. We completed a factorial ablation and added comparative experiments. We respond to each question below.

Q1: Can the effects of the integrator, Ampère loss, and initial-derivative constraint be examined separately?

We conducted a factorial study with two integrators, two independently switched constraints, and five random seeds, for 40 runs in total. All runs used the same network, sampling, and budget of 3600 updates. The setting was Landau damping (T=8, α=0.1, K=4) with a 4×128 network. The table reports mean terminal relative distribution error (%) over five seeds.

IntegratorFull modelWithout AmpèreWithout initial derivativeWithout both
Symmetric0.1630.2940.1920.444
RK40.1740.2740.1890.454

Adding the Ampère term lowers mean error for both integrators. With Ampère retained, the initial-derivative term reduces the symmetric-integrator distribution error from 0.192% to 0.163% and the initial residual RMS from 1.52×10⁻² to 7.37×10⁻⁶. The full model performs similarly with both integrators.

Q2: What are the training windows, sampling strategy, velocity quadrature, values of M_s, and treatment of trajectories leaving the velocity interval?

Thank you for pointing out the details needed for reproducibility. The original settings are:

SettingGaussianLandau
Covered time interval[0,1][0,2]→[0,4]→[0,6]→[0,8]
Updates per stage2000800/800/800/1200
Time slices8, aligned to 0.01Aligned to 0.1; half old and half new interval
Spatial points per slice64 fixed uniform points64 fixed uniform points
Velocity quadrature[-6,6), 96 cell centers[-6,6), 128 cell centers
M_s10020/40/60/80

Landau training progressively expands the time interval. After the first stage, half of the slices revisit the covered interval and half sample the newly added interval. The original run used four slices for the first 1600 updates and eight thereafter; the Q1 ablations used eight throughout.

Velocity integrals are evaluated at equal-width cell centers, weighted by cell width. The smoothing term uses M_s=round(T_s/h), where h is the time step, T_s is the length of the covered interval [0,T_s] (2, 4, 6, or 8), and M_s is the number of smoothing intervals; round denotes rounding to the nearest integer. The smoothing weight is 10⁻⁴. During backtracing, only position is mapped periodically. Velocity is neither clipped nor mapped periodically; even outside [-6,6), the analytic initial condition is evaluated at the backtraced point (x₀,v₀).

Q3: What does the neural field representation add relative to NuFI?

Thank you for this important comparison. For tasks requiring repeated queries of distributions at different positions and times, NFCR needs only one model covering the full time interval and reconstructs each requested distribution on demand, without pre-storing many full snapshots; adding query times or positions does not increase model storage. After training, NFCR directly generates the complete terminal distribution and electric field, with substantially faster output than NuFI and SL solving from the initial state.

All three methods use the same initial condition, 128² evaluation grid, and reference solution. NFCR uses a two-layer, width-32 network (1448 parameters), 3600 Adam updates, and L-BFGS refinement. Solution storage comprises NFCR parameters and buffers, NuFI field history and fixed data, and 161 full SL distribution snapshots.

MetricNFCRNuFICubic-spline SL
Maximum relative mass drift1.37×10⁻⁸2.08×10⁻⁸5.85×10⁻⁵
Distribution error over the trajectory0.04778%0.00562%0.02024%
Terminal distribution error0.06726%0.01120%0.03362%
Solution storage over the trajectory5856 B5168 B10.0625 MiB
Time to produce one terminal state0.0611 s5.6205 s2.1792 s
Peak GPU memory for one terminal state1.215 MiB1.048 MiB7.021 MiB

NFCR keeps mass drift at the 10⁻⁸ level. One model of about 5.72 KiB represents the full interval; adding query times or positions does not increase its storage. Its size is comparable to NuFI and about 1802 times smaller than 161 SL snapshots. NFCR generates the complete terminal distribution and field in 0.0611 seconds, about 92.0 and 35.7 times faster than NuFI and SL solving from the initial state, respectively, with 82.7% lower peak memory than SL. Times are medians of five runs after warm-up. Offline training and refinement take 31.05 minutes.

Thank you again for helping us clarify the component roles, reproducibility settings, and properties of the neural field representation.

The detailed responses and additional experiments have largely addressed my main concerns, so I am raising my score by one point.

Reviewer_us7LOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes NFCR, a method for the 1D-1V Vlasov-Poisson system in which a neural network represents only the electric field. A small network of time outputs the first K spatial Fourier coefficients of the field, anchored at the initial field (Eq. 5). The distribution at any query point is obtained by tracing characteristics backward to t = 0 under this field with a symmetric drift-kick-drift splitting and evaluating the known initial condition (Eqs. 6-7). The network is trained without labels by penalizing Poisson and Ampère residuals computed from the velocity moments of the reconstructed distribution, together with an initial-derivative condition and a smoothness regularizer, over progressively longer time windows. The authors note that the integrator is locally area preserving and that the reconstruction preserves nonnegativity. Experiments on a weakly coupled Gaussian transport problem (t ≤ 1) and linear Landau damping (t ≤ 8) compare against a semi-Lagrangian solver, particle-in-cell, a PINN, and a semi-Lagrangian PINN, and report lower distribution and field errors against a high-resolution classical reference, as well as small mass drift.

Critical strengths

The basic design choice is sensible: learning only the comparatively smooth, low-dimensional electric field and generating the fine-scale structure of the distribution by characteristic transport from the known initial condition is a better use of a network than fitting the filamenting distribution directly, and it gives nonnegativity and local phase-space area preservation for free. Against the other neural approaches considered (PINN, SL-PINN), the gains are large.

Critical weaknesses
  1. The paper does not motivate why we want to solve this problem, or why a neural method is needed to do it. The introduction argues that Vlasov-Poisson is hard because the distribution develops fine structure, but never says why one wants to solve it: there is no discussion of applications or of the technological implications of a better solver (plasma physics is never mentioned). Nor does it describe the state of the art in classical numerical methods or explain what limitation of those methods NFCR is meant to address. More generally, the introduction and background (Secs. 1-2) read as a dense list of existing techniques (semi-Lagrangian methods, PINNs, SL-PINNs, NuFI) without the context needed to understand what they are or why we should care about them, which will be especially hard to follow for an ICLR audience.
  2. The experiments do not show that the method adds anything over classical numerics. With proposals like this, the burden is on the ML method to demonstrate that it can solve an interesting problem that could not be solved by other means. Both test cases are 1D-1V, i.e., a two-dimensional phase space, which is routinely and cheaply solved by classical grid-based (semi-Lagrangian, Eulerian, spectral) or particle methods; indeed, the reference solution used to score every method is itself a classical solver on a 1024 × 1024 grid. The problems are also easy within this class: the Gaussian case is weakly coupled (κ = 0.1) with a field of amplitude around 0.04 over t ≤ 1, so it is close to free streaming, and the Landau case is linear damping (amplitude 0.1) up to t = 8 with an initial field that already lies in the K = 4 mode basis. Standard harder benchmarks such as nonlinear Landau damping, two-stream instability, or longer time horizons are absent, as is any higher-dimensional phase space, which is the regime where classical methods actually become expensive. The classical baselines are also weakly configured: the semi-Lagrangian solver uses linear interpolation with Δt = 0.05 (App. A.2) rather than the standard cubic splines, PIC is scored on an unsmoothed histogram so its error is dominated by sampling noise, and NuFI, the closest classical method and the one this work builds on, is not compared. No wall-clock or cost comparison is given. Finally, the headline Landau field error (2.7 × 10⁻⁶, Table 2) is from a single seed, while the five-seed mean in App. D is 8.8 × 10⁻⁵ ± 1.6 × 10⁻⁴. If the problems are simple and the method essentially matches a classical reference, I don't see what it adds at the end of the day.
Questions for authors
  1. (Clarification) What is the intended application of the method, and in what setting do the authors expect it to be preferable to established classical solvers?
  2. (Experiments) Can the authors compare against a well-configured classical solver (e.g., semi-Lagrangian with cubic splines, and NuFI) at matched wall-clock cost, and report the cost of NFCR training and querying?
  3. (Experiments) How does the method perform on harder problems, such as nonlinear Landau damping or the two-stream instability over long times, and on a higher-dimensional (e.g., 2D-2V) phase space where classical grid-based methods become expensive? Results in such a regime would be needed to change my assessment.
1 comment

Thank you for reviewing our paper and noting the need to clarify its intended use, compare it with classical methods, and add higher-dimensional validation. We have added relevant experiments and respond to each question below.

Q1: What is the intended application? In what settings might the method be useful?

Compact storage and fast on-demand reconstruction. For tasks that require repeated access to distributions at different positions and times, NFCR stores one model for the full time interval and reconstructs each requested distribution on demand, instead of storing many full snapshots in advance. Adding query times or positions does not increase model storage. In 1D1V Landau damping, the trained NFCR model generates the complete terminal distribution and field in 0.0611 seconds, about 92.0 and 35.7 times faster than NuFI and SL solving from the initial state. Q2 gives the corresponding accuracy and resource results; Q3 shows how this representation performs in six-dimensional phase space.

Q2: How does NFCR compare with NuFI and a strong classical method in accuracy and cost?

We compare NFCR, NuFI, and a semi-Lagrangian (SL) method with cubic-spline interpolation on 1D1V Landau damping (T=8), using the same initial condition, 128² evaluation grid, and reference solution. NFCR uses a two-layer, width-32 network, 3600 Adam updates, and L-BFGS refinement, without reference-solution labels.

MetricNFCRNuFICubic-spline SL
Maximum relative mass drift1.37×10⁻⁸2.08×10⁻⁸5.85×10⁻⁵
Distribution error over the trajectory0.04778%0.00562%0.02024%
Terminal distribution error0.06726%0.01120%0.03362%
Solution storage over the trajectory5856 B5168 B10.0625 MiB (161 snapshots)
Time to produce one terminal state0.0611 s5.6205 s2.1792 s
Peak GPU memory for one terminal state1.215 MiB1.048 MiB7.021 MiB

NFCR keeps mass drift at the 10⁻⁸ level. One model of about 5.72 KiB represents the full evolution interval; adding query times or positions does not increase its storage. Its size is comparable to NuFI and about 1802 times smaller than 161 SL snapshots. NFCR generates the complete terminal distribution and field in 0.0611 seconds, about 92.0 and 35.7 times faster than NuFI and SL solving from the initial state, respectively; its peak memory for this output is 82.7% lower than SL's. Times are medians of five runs after warm-up. Offline training and refinement take 31.05 minutes, and the trained model remains available for later queries.

Q3: How does the method perform for higher-dimensional or more complex dynamics?

We test six-dimensional distribution reconstruction on 3D3V Landau damping. All three methods use a 4³×24³ grid and time step 0.05; NFCR uses a two-layer, width-32 field network. The table reports errors over nine time points, storage for the full interval, and the cost of generating one complete terminal distribution and field.

MetricNFCRNuFI (3D extension)Cubic-spline SL
Full-distribution error0.008362%0.007970%0.1220%
Perturbation-distribution error0.4861%0.4633%7.0908%
Field error0.2579%0.1212%7.0324%
Solution storage8428 B8580 B30.375 MiB (nine snapshots)
Time to produce one terminal state0.2624 s9.9323 s4.5583 s
Peak GPU memory for one terminal state40.465 MiB50.701 MiB313.508 MiB

In this six-dimensional case, NFCR's perturbation error is close to NuFI's and lower than the SL result in this comparison. Its model storage is comparable to NuFI's and much smaller than the SL snapshot history. NFCR generates one terminal state about 38 times faster than NuFI and 17 times faster than SL, with lower peak GPU memory than either method. Times are medians of five runs after warm-up. These results show that a small field network can represent evolution in a high-dimensional phase space and enable fast terminal-state reconstruction after training.

Multi-seed results and clarification of the field error

Thank you for raising the representativeness of a single-run result. The terminal field error of 2.687×10⁻⁶ in Table 2 is the lowest among six runs. That run is not the best in terminal distribution error or trajectory-wide field error, so it should not be taken as representative field accuracy.

Under the same evaluation protocol, the five-seed mean terminal field error is 8.793×10⁻⁵, with a sample standard deviation of 1.583×10⁻⁴ and a median of 2.714×10⁻⁵, indicating seed sensitivity in this metric. We will identify the Table 2 value as a single-run result and add per-seed results and the error evolution over time.

Thank you again for helping us clarify the intended use of the method and its performance in higher-dimensional reconstruction.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.