- 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
- 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.
- 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
- 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.
- 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.
- 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.
- 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
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:
| Model | before → after (dB) | Flux relL1 before → after | before → after (dB) |
|---|---|---|---|
| AE | 15.42 → 26.57 | 0.131 → 0.032 | 35.49 → 35.88 |
| VAE | 9.33 → 24.18 | 0.193 → 0.063 | 34.35 → 34.75 |
| VQ-VAE decoder, continuous code | 12.91 → 22.97 | 0.218 → 0.087 | 34.16 → 34.50 |
| GyroSplat | 27.69 | 5.2e-7 | 37.63 |
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:
| Check | CR 302 | CR 142 |
|---|---|---|
| Flux relL1 below 1e-4 | 39/39 | 39/39 |
| Net heat-flux sign flips | 0/39 | 0/39 |
| κ_rel within 10% of the reference | 36/39 | 38/39 |
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.
| Method | CR (byte-wise) | Encode (s) | Decode (s) | Peak GPU memory (GB) |
|---|---|---|---|---|
| GyroSplat | 302 | 169 | 0.042 | 19.0 |
| NF + physics | 302 | 85 | 0.166 | 55.0 |
| GyroSplat | 142 | 399 | 0.088 | 39.1 |
| NF + physics | 142 | 116 | 0.245 | 94.2 |
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:
| Stage | CR 302 (s) | CR 142 (s) |
|---|---|---|
| Distribution fit | 87 | 184 |
| Potential fit | 6 | 10 |
| Gabor + joint fit | 64 | 182 |
| Final Gauss–Newton refinement | 11 | 24 |
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.
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:
[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). We now also evaluated it by radial wavenumber, Q(kx), by position along the field line, Q(s), and by velocity, 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):
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 ky 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.