acceptodds
Under review as a conference paper at ICLR 2027

GyroSplat: Exact and Differentiable Linear Representations of Plasma Turbulence

Abstract

Gyrokinetic simulations model the plasma turbulence governing heat transport in magnetically confined fusion devices. Due to their 5D nature, these simulations are computationally expensive and produce massive data volumes, with a single snapshot requiring tens of gigabytes of storage. This makes directly storing intermediate timesteps impractical and motivates the use of (lossy) compression. Recent neural compression approaches achieve favorable rate–distortion trade-offs by augmenting reconstruction objectives with physics-informed losses that preserve derived quantities such as the electrostatic potential and heat flux. However, neural network based representations can be difficult to optimize and refine when additional constraints are imposed. We introduce GyroSplat, a compression method for gyrokinetic fields inspired by Gaussian splatting. GyroSplat represents each plasma field as a weighted sum of independent Gaussian–Gabor splats, yielding a representation that is linear in its coefficients. This linearity enables a staged optimization strategy in which splats are first fitted to minimize reconstruction error, then refined against the potential, and subsequently new splats are added to correct the transport. Moreover, derived quantities admit closed-form expressions in terms of the splat coefficients, enabling efficient refinement via a Gauss–Newton procedure. Experiments on gyrokinetic simulations show that GyroSplat achieves improved trade-offs at matched storage over established compression baselines on derived physical quantities while remaining competitive on field reconstruction. These results highlight Gaussian splatting as an effective and flexible approach to high-fidelity compression of scientific simulation data.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
scientific data compression, Gaussian splatting, gyrokinetics, plasma turbulence, physics-informed compression, Gabor atoms, Gauss–Newton optimization, nuclear fusion

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_HsddOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces GyroSplat, a five-dimensional Gaussian-Gabor representation for gyrokinetic plasma snapshots. The staged fitting procedure is technically motivated and produces strong results on the reported dataset. However, the main physical metric is optimized on each snapshot and evaluated on the same heat-flux modes, the storage comparison is not a common serialized-byte comparison, and the learned baselines receive asymmetric physics supervision.

Critical strengths
  1. The paper addresses a meaningful problem in gyrokinetic simulation, where five-dimensional snapshots are expensive to store while derived quantities such as electrostatic potential and heat flux remain scientifically important.
  2. The three-stage fitting procedure is clearly motivated. The roles of base splats, Gabor splats, and fixed-geometry amplitude refinement are easy to follow, and Figure 1 gives a useful overview.
Critical weaknesses
  1. The strongest physical result is not independently validated. Stage 3 directly optimizes the heat-flux spectrum of each snapshot over all nonzero ky modes 1-31, and the same modes are then used for evaluation. The reported 10^-7 to 10^-8 relative errors therefore show that the target diagnostic can be fitted, but they do not establish preservation of unseen diagnostics, future time slices, or downstream simulation behavior. The authors should hold out ky modes, time slices, or trajectories and report particle flux, momentum flux, or restart tests.
  2. The learned baselines are not trained with the same physical supervision. GyroSplat directly uses potential and heat-flux objectives, whereas the AE, VAE, and VQ-VAE rows are mainly reconstruction-only. The comparison therefore mixes representation quality with supervision and optimization budget. A fair comparison requires physics-aware versions of these baselines or a narrower claim.
  3. The fitting pipeline is non-convex and uses a single fixed random seed. The paper does not report seed variation, failure rates, damping sensitivity, or the full definitions of several optimization hyperparameters. Multiple seeds and complete configurations are needed to assess stability and reproducibility.
  4. The practical efficiency claim is not quantified. Encoding time, decoding time, operator construction time, peak memory, and cache cost are not reported, although the method uses several optimization stages and repeated diagnostic-operator updates.
Questions for authors

The paper contains a promising and reasonably well-motivated combination of Gaussian-Gabor splats and physics-aware coefficient refinement. The current experiments support a potentially useful task-specific compression method, but the strongest claims would benefit from independent physical validation, byte-level storage accounting, and matched physics-aware learned baselines. With these clarifications or a narrower statement of scope, the work could make a solid contribution.

3 comments
AuthorsOpenReview ↗

Evaluation and further experiments

We thank the reviewer for raising this interesting point. At its core, it asks whether a compressor can validly be evaluated on the quantities its objective targets. Goodhart’s law warns that “when a measure becomes a target, it ceases to be a good measure.” We argue that this does not apply here, and we back this with evidence beyond the objective.

Compression requires choosing what to keep. By the pigeonhole principle, no lossy compressor can keep all information, so deciding what to keep is part of the design. JPEG, for example, discards high frequencies according to human visual thresholds [1]. For compressors that rely on learning or optimization (e.g. with SGD), this choice is usually shaped largely by the loss. As also the reviewer mentions, a quantity outside the objective is therefore not guaranteed to be preserved. Optimizing for the heat flux spectrum is thus part of the compression design, not circular evaluation. Goodhart’s law would apply only if we treated the spectrum error as evidence of overall fidelity. Instead, we assess overall fidelity through PSNR and quantities that are not optimized (see below).

The concerns are addressed in our evaluation and further experiments:

  • Held-out test set: All hyperparameters were chosen on training and validation snapshots, and all results are reported on a held-out test set. Holding out time slices from the compressed data, as suggested, is not meaningful, because a compressor is by definition evaluated on the data it compresses.
  • Non-optimized quantity κ\kappa: Our original evaluation already includes κ\kappa, which is not directly optimized.
  • Additional experiments: Following the reviewer’s suggestion, the revised paper adds fluxes from other dimensions for evaluation, which have not been refined.

[1]: CCITT, “Information technology – Digital compression and coding of continuous-tone still images – Requirements and guidelines,” ITU-T Recommendation T.81, Sep. 1992, Annex K.1.

Heat flux in other dimensions Our objective refines the heat flux as a function of the binormal wavenumber, Q(ky)Q(k_y). We now also evaluated it by radial wavenumber, Q(kx)Q(k_x), by position along the field line, Q(s)Q(s), and by velocity, Q(v∥,μ)Q(v∥, μ) (parallel velocity and magnetic moment). Shown below are the median relative L1 errors, of 39 snapshots at a compression ratio (CR) of 302 (byte-wise) (see Appendix G, Table 15):

MethodQ(kx)Q(k_x)Q(s)Q(s)Q(v∥,μ)Q(v∥, μ)
GyroSplat0.1420.0720.433
NF + physics0.1590.0440.192
Fourier + refinement0.1460.1200.758
Autoencoder0.1060.0690.108

In the spatial views GyroSplat is close to the other methods. In velocity space, NF + physics and the autoencoder are more accurate.

Held-out modes On nine snapshots, we fitted only the even or only the odd kyk_y modes. The fitted modes stay at 1e-8 to 1e-7 error, the held-out modes have 21% and 33% error, and the potential PSNR changes by less than 0.1 dB (Table 16). The exact flux holds only for the fitted modes.

Solver restarts We restarted a gyrokinetic solver for eight snapshots and their reconstructions at the compression ratio (CR) of 302 (the restart from the original is the reference). Over the first five simulation time units, heat-flux errors are 2.4%, 3.6% and 12.2% for GyroSplat, the autoencoder and NF + physics; flux first departs by 10% after 22, 15 and 7 time units, versus 72 for a 1e-6 perturbation of the original (Table 17).

What we claim The results show that fine-tuning of chosen diagnostics at our reported field and potential accuracy work well and are accurate, but this does not predict the preservation of other quantity. We revised the paper to state this more clearly.

AuthorsOpenReview ↗

AE refinement, reproducibility, timing and memory cost

AE supervision

Physics refinement for autoencoders This is a fair concern! The reconstruction-only autoencoder (AE) rows do not show the difference between optimizing the representation and the optimization of observable quanties. Therefore, we have added physics refinement for AE, VAE and VQ-VAE latents. PINC reports unstable end-to-end AE training with physics refinment, and we also encountered difficulties with adapter fine-tuning. To this end, we kept the pretrained decoders fixed/frozen and optimized their stored latents/codes, similar to the NF + physics and Fourier/Haar comparisons at matched byte budgets.

Latent refinement experiment The Table below shows what changes after 400 Adam steps with field, potential and flux losses. Values are medians over all 39 snapshots at approximately 300 CR:

ModelPSNRφPSNR_φ before → after (dB)Flux relL1 before → afterPSNRfPSNR_f before → after (dB)
AE15.42 → 26.570.131 → 0.03235.49 → 35.88
VAE9.33 → 24.180.193 → 0.06334.35 → 34.75
VQ-VAE decoder, continuous code12.91 → 22.970.218 → 0.08734.16 → 34.50
GyroSplat27.695.2e-737.63

The remaining reconstruction trade-off The additional losses improve all three metrics for each model, and the AE becomes competitive on potential accuracy. When we try to push the AE flux fit to numerical precision: potential PSNR falls to about 10 dB. Comparing this to GyroSplat, they keep the higher field and potential accuracy at a similarly small flux error. Appendix H gives the combined results in Table 18 and the longer, 2,000-step fits in Table 19, so readers can also see what changes with more fitting steps.

Stability and reproducibility

Multiple seeds and stability checks We have added the fitting on multiple seeds, stability checks and fitting details you asked for. Appendix D reports statistics on multiple seeds, and Appendix J checks all 39 compressed fits at each byte-wise budget:

CheckCR 302CR 142
Flux relL1 below 1e-439/3939/39
Net heat-flux sign flips0/390/39
κ_rel within 10% of the reference36/3938/39

Damping and difficult cases We also varied damping and step count on 12 snapshots, including the hardest snapshots. Damping from 1e-8 to 1e-4 gives small flux residuals with little median change in potential PSNR. The hardest case still loses about 1 dB whenever its flux is fitted. Changing damping did not remove that trade-off in this case. Appendix J (Tables 22–23) reports the checks and convergence results.

Complete fitting details To make the fits reproducible, Appendix I (Tables 20–21) now gives the optimizer settings, losses, normalization, residual scaling, damping and adaptive branch.

Timing and memory cost

Timing and memory measurements We have added encoding and decoding times, peak memory and a breakdown of the fitting stages. Methods ran in turn on dedicated B300 GPUs, with one job per GPU. The table gives medians over nine pilot snapshots and two seeds, after excluding warm-up and shared-GPU runs. Encoding includes the complete fit of a snapshot, while decoding reconstructs/renders the full 5D field.

MethodCR (byte-wise)Encode (s)Decode (s)Peak GPU memory (GB)
GyroSplat3021690.04219.0
NF + physics302850.16655.0
GyroSplat1423990.08839.1
NF + physics1421160.24594.2

Time spent in each stage GyroSplat takes longer to compress a snapshot, while decoding is faster and peak GPU memory is lower. The stage breakdown shows where the time is spent during fitting:

StageCR 302 (s)CR 142 (s)
Distribution fit87184
Potential fit610
Gabor + joint fit64182
Final Gauss–Newton refinement1124

Readout setup and fitting memory Preparing simulation geometry and reference readouts takes about 2.5 s. During fitting/encoding, the code holds about 223 MB of reference arrays, 3.6 MB of shared trajectory geometry and a 1.8/3.9 MB Jacobian. The reference arrays and Jacobian are discarded after fitting.

Full cost measurements Appendix K (Tables 24–25) gives the timing distributions, memory use and stage costs. Appendix E reports the shared decoder storage separately from the cost of compressing each snapshot.

Thanks for your detailed response. It addresses some of my concerns. I will maintain my positive score.

Reviewer_NdGSOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose GyroSplat, a compression method for 5D gyrokinetic data in plasma physics. The method consists of three stages: (1) fitting basic Gaussian splats; (2) adding Gabor splats and refine the fit; (3) modulating Gabor weights.

Critical strengths
  1. The size of gyrokinetic data is a known problem. To the best of my knowledge, I have not found earlier work using Gaussian splat for gyrokinetic data compression. So the proposed method looks novel.
  2. The results are considerably better compared to baselines.
Critical weaknesses
  1. The target research gap is not well-defined. In introduction, the authors claim "existing approaches typically struggle to compress field data effectively while preserving derived quantities of interest". Here, "effectively" is vague. It would be clearer if the authors can describe how much compression ratio is needed, with how much tolerance of fidelity loss.
Questions for authors
  1. Qualitatively, what research problem is addressed? In other words, this work achieves lower reconstruction error under 142x and 302x compression rate, compared to baselines -- why is the improvement significant? It is an interesting paper, and it can be better if the motivation is deeper than having better numerical results.
2 comments
AuthorsOpenReview ↗

Research gap and why improvement matters

W1 — Research gap?

The research gap is compression when optimizing for general L2 reconstruction and derived quantitites. To show the effectiveness of the method, one could

  • (a) either go for fixing a quality target for a specific use-case (e.g. visualization or model training) and then optimize for it.
  • (b) show the more general quality-compression tradeoff by evaluating several points.

We opted for (b) as it should be useful to a broader audience than (a). In such an analysis it is computationally cheaper to fix compression ratios and then compare the resulting quality. Our method is effective in the sense that it improves quality on the fixed compression budgets we evaluated.

Note that as part of the rebuttal we extended GyroSplats to different compression ration ranging from 100 to 1000 (see reponse to reviewer GLVf).

Q1 — Why does the improvement matter?

The research problem we address is how to compress high-dimensional data while preserving both the L2 reconstruction and specific derived quantities. Our results are significant because the Splat representation lets us evaluate these derived quantities in closed form, so we can optimize them directly. The results on derived quantities such as heat flux show that this optimization reaches essentially numerical precision.

Thank you for your rebuttal. My concerns are addressed and I will maintain the score.

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

GyroSplat compresses 5D gyrokinetic data into a weighted sum of Gaussian–Gabor splats. GyroSplat uses a three-stage pipeline: (1) fit the field, (2) fit the potential, and (3) fit flux, and ending with a Gauss–Newton refinement of the Gabor amplitudes at fixed geometry. The authors test on 39 electrostatic snapshots at ~302× and ~142× compression, demonstrating best potential PSNR and near-zero flux error, while 5D field PSNR is comparable to the best baselines.

Critical strengths
  1. This is a well-motivated method. Linear potential and quadratic flux in the coefficients gives closed-form Jacobians and an efficient Gauss–Newton step.

  2. The authors propose a useful new metric ( κ/κ_truth). Results shows that low net-flux error and good PSNR can hide badly wrong outward/inward transport, motivating this metric.

  3. The study has good baselines and ablations. Baselines cover classical codecs (several wavelet bases, PCA, ZFP, SZ3, Fourier), learned codecs, and physics-refined neural fields and transform codes. The appendix reports IQRs, per-regime medians, and sign-flip counts.

Critical weaknesses
  1. The paper is a bit limited in the scenarios evaluated. It only considers a subset of electron data, with 39 snapshots from 27 trajectories. The paper is, however, motivated more broadly. As such, the paper feels more suited to a venue more specific to this data regime than a general learning conference.

  2. The paper could be more thorough in terms of statistical reporting. There is a single fixed seed, no variation across fits, and median-only headline numbers on 39 snapshots.

  3. The evaluation with respect to compression baselines and codecs is quite limited, which is a major weakness since the paper's motivation is largely driven by compression arguments. For example, there should be a more direct evaluation in terms of bytes used. In addition, a more thorough tradeoff curve in terms of compression to reconstruction performance is needed rather than just evaluations at two arbitrary compression factors (e.g., 302x).

Questions for authors
  1. Results come from a single fixed seed and are reported as medians. Can you report variance across fitting seeds (e.g., mean ± std over several seeds on a subset of snapshots) to show the gains over the baselines are not within seed noise?

  2. Compression ratios are counted in stored reals and only at two points (302× and 142×). Can you report serialized bytes (including support maps, metadata, and shared decoders) and show full rate–distortion curves across a range of CRs, for both reconstruction and physics metrics?

1 comment
AuthorsOpenReview ↗

Scope, multiple seeds, byte-wise and rate-distortion

W1 — Scope

Compressing a high-dimensional field while preserving quantities derived from it is a general and important problem which justifies the submission to this venue. Gyrokinetics is a demanding instance, with five-dimensional ion distribution functions of 22.3 million values per snapshot and electrons treated adiabatically. Because the splat representation is linear in its coefficients, any linear or quadratic diagnostic becomes closed-form and differentiable, so the approach is not tied to the heat flux. Our experiments cover one setting, and we expect these lessons to carry over to other high-dimensional simulation data, which we leave to future work.

W2 / Q1 — Muliple seeds

Muliple seeds To check whether the gains depend on the inital results, we repeated GyroSplat and (Neural Fields) NF + physics with five seeds on all 39 snapshots at both budgets. For each seed we take the median over snapshots. The table reports the mean and standard deviation of those medians.

MethodPSNR_f (dB)PSNR_φ (dB)Flux relL1
GyroSplat, 302×37.63 ± 0.0227.69 ± 0.045.2e-7 ± 2.0e-7
NF + physics, 302×37.29 ± 0.1026.63 ± 0.200.142 ± 0.002
GyroSplat, 142×38.21 ± 0.0329.24 ± 0.063.7e-8 ± 0.5e-8
NF + physics, 142×38.16 ± 0.1427.56 ± 0.250.105 ± 0.003

GyroSplat continues to reconstruct the potential more accurately on multiple seeds. After averaging the seeds for each snapshot, it is better than NF + physics on 35/39 snapshots at CR 302 and 38/39 at CR 142. Appendix D, Table 7 gives the full seed results, with the statistical comparisons.

W3 / Q2 — Bytes and rate–distortion

Comparison in bytes We agree that byte comparision and a wider rate–distortion comparison are needed to evaluate and compare our compression method. Tables 1 and 3 and the comparison figures now use byte-wise budgets throughout. The main conclusions drawn remain unchanged. At approximately 302 CR, the results over all 39 snapshots are (see Table above).

MethodPSNR_f (dB)PSNR_φ (dB)Flux relL1
GyroSplat37.6327.695.2e-7
NF + physics37.2926.630.142
Splats, field loss only39.327.380.271
Fourier + refinement35.2323.190.024
Fourier top-K35.2521.360.289
Haar top-K34.40−8.321.01
bior4.4 top-K34.40−12.871.14

Rate–distortion curves To address the concern about the two original compression budgets, Figure 3 now covers eight byte-based compression ratios from 1000 to 100, with the full comparison in Appendix F. From CR 600 to 100, GyroSplat has the highest potential PSNR and the lowest flux error. At CR 800 and 1000, NF + physics reaches a slightly higher potential PSNR, while GyroSplat keeps the lower flux error. We attribute this at least in part to tuning. GyroSplat’s hyperparameters were selected at CR 142 and 302 and were not re-tuned for higher compression ratios. This includes the fraction of Gabor splats, which sets how much of the budget goes to oscillatory structure and is currently not learned from the data. We hypothesize a neural field may adapt more easily as the budget shrinks, because it learns the frequency content to represent rather than having it set in advance. Choosing the Gabor fraction automatically is a natural extension.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

Additional experiments We thank the reviewers for reading the paper and for the specific suggestions. We have repeated the fits with several seeds, compared the methods in bytes, and added rate–distortion curves across a wider range of budgets. We also applied physics losses to the autoencoder codes and measured encoding time, decoding time and memory use. These additions are included in the revised paper. The changes made to the paper are highlighted in blue.

What the results establish We have also clarified what the small heat-flux errors mean. Since each snapshot is fitted to its own target diagnostics, those errors show how closely the fit matches the targets. We tested other heat-flux views, held-out modes and early solver restarts to see what else is preserved. The results are report in the appendix and limit our claims accordingly.

One major change is that we now report all compression ratios based on the suggested byte-wise comparison. We think this fits the paper better; another thanks for the suggestions of the reviewers!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.