- Rating
- 4(+1)
- Confidence
- 4
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
LATEC is a learned compression method for scientific data with varying dimensionality. It combines neural prediction with classical compression techniques to improve compression efficiency while maintaining pointwise error bounds. A shared model trained on synthetic data can be applied across different dimensionalities and scientific domains without retraining. Evaluation on diverse scientific datasets shows competitive compression performance, with further gains from domain-specific fine-tuning.
Critical strengths
Practical relevance and generality. I see scientific data compression as an important problem that deserves more attention from the deep learning community. LATEC uses a shared architecture across dimensionalities, while many existing compression frameworks that incorporate deep learning assume a fixed input dimensionality. This makes it easier to apply to different scientific datasets without redesigning the model. I see this work as a useful starting point for encouraging further research in this practically important area.
Synthetic pretraining and broad evaluation. I also like that a model trained on synthetic data works well on real scientific datasets, including unseen dimensionalities, without target-domain fine-tuning. I particularly appreciate the evaluation across diverse datasets and several established compressors. Bringing these comparisons together takes considerable effort and provides stronger evidence of the method’s generality and competitive performance under pointwise error constraints.
Critical weaknesses
Limited practical efficiency. The compression gains come with a substantial computational cost: LATEC is considerably slower than SZ3 for both encoding and decoding. Given that scientific compression often aims to reduce I/O costs, I would like to better understand when the bitrate savings outweigh this overhead. This limits the practical benefit demonstrated by the current results.
Modest gains on real high-dimensional data. While the shared model handles unseen dimensionalities, its gains over SZ3 on the two real 5D datasets are limited. This raises the question of whether training a model specifically for the input data at compression time could yield larger compression gains. The throughput implications are unclear, however, so comparing both compression performance and total encoding time would help clarify this trade-off.
Questions for authors
-
Synthetic pretraining and data-specific adaptation. Could the limited gains on the real 5D datasets reflect a mismatch between the synthetic training distribution and the structure of these fields? How much does domain-specific fine-tuning close this gap? If the gains remain limited, could training a model specifically for the input data at compression time help? A comparison or discussion of approaches such as NeurLZ, considering both compression performance and throughput including training time, would help clarify this trade-off.
-
Practical efficiency. In which use cases would LATEC’s bitrate savings outweigh its substantially lower encoding and decoding throughput? An estimate or an end-to-end example would help clarify when it would be preferable to SZ3.
-
Benefit of gating. Figure E.5 shows how often each predictor is selected. Could you also clarify how much bitrate reduction scale-guided gating provides compared with using the GNN alone?
-
Model capacity and efficiency. The reported results use a fixed GNN architecture. Have you explored how model size affects compression performance and throughput? In particular, could a smaller model retain most of the compression gains while reducing the encoding and decoding cost?
4 comments
Thank you for the detailed response and the additional experiments. I appreciated that the per-input fine-tuning experiment accounted for both training time and the cost of storing weight updates.
I’d still be interested in a direct comparison with NeurLZ. It trains a small residual model for each input block, which can keep both training and inference fast. The NeurLZ paper reports compression gains, and I’m curious how those gains compare with LATEC at the same error bound. Could you compare the two on the same data, reporting the final bitrate including model weights, total encoding time including training, and decoding time?
The checkpointing example addressed my concern about LATEC’s lower throughput relative to SZ3. When checkpoints are rarely read back, smaller files can matter more than faster decoding. The other experiments addressed my remaining concerns. Thank you again.
We thank the reviewer for their feedback. To answer your question, we also compared LATEC with NeurLZ. On average, LATEC saves 40.4% vs 35.9% for NeurLZ. We did not find any open source implementation of the algorithm, so we follow NeurLZ's own protocol on the same data to be able to reuse the reported results in the paper (table 2): the same NYX, Miranda, and Hurricane fields, and the same error bounds (1E-2, 5E-3, 1E-3, 5E-4, 1E-4). We compute the bit-rate saving over SZ3 v3.1.8 (while our benchmark used v3.4) at equal PSNR, averaged over the 5 bounds. Here are the detailed results:
| Method | Nyx Temperature | Nyx Dark Matter Density | Nyx Velocity Y | Nyx Baryon Density | Miranda Diffusivity | Miranda Velocity Z | Miranda Viscocity | Hurricane Cloud | Hurricane Precip | Hurricane W | Mean |
|---|---|---|---|---|---|---|---|---|---|---|---|
| NeurLZ | 41.84 | 67.92 | 29.53 | 75.66 | 31.72 | 26.88 | 31.12 | 11.71 | 18.69 | 23.38 | 35.85 |
| LATEC | 43.84 | 70.79 | 33.10 | 43.97 | 15.13 | 26.10 | 16.52 | 50.00 | 57.44 | 47.05 | 40.39 |
NeurLZ encoding times are derived from the ratios in its Table 3 (Nyx 512^3), applied to our SZ3 times. Training is extrapolated linearly from 10 to 100 epochs. Miranda and Hurricane are not in Table 3, and the ratios grow for smaller fields, so those estimates are likely low.
| Dataset | SZ3 | NeurLZ (est.) | LATEC |
|---|---|---|---|
| Encode | |||
| Nyx | 3.8 | ~360 | 6.7 |
| Miranda | 1.18 | ~110 | 2.2 |
| Hurricane | 0.64 | ~60 | 1.9 |
| Decode | |||
| Nyx | 1.6 | ~3.0 | 4.8 |
| Miranda | 0.53 | ~1.0 | 1.4 |
| Hurricane | 0.31 | ~0.6 | 1.0 |
| Hardware | |||
| Device | CPU | 8x RTX 6000 Ada | 1x RTX 5070 |
Note that NeurLZ timings are reported from their setup, using 8 GPUs while LATEC only uses 1.
Thank you for the additional experiments. The higher encoding cost of online training makes sense and matches my intuition. LATEC’s bitrate savings are impressive, and I’ve raised my rating to 4.
We are grateful to the reviewer for the encouraging assessment and for very precise questions. They led to the new gating, model-width, per-input experiments and the scale shift improvement.
W1, Q2. Practical efficiency. With the Triton and Numba implementation (see W1 in the global response), the gap to SZ3 is much smaller than the 12-24× of the submission (Sec. 3, Fig. 9, App. E.4). On one H100, LATEC encodes at 220-270 MiB/s from 2-D to 4-D and 153 MiB/s in 5-D. SZ3 on 8 CPU threads encodes 1.4-1.5× faster in 3-D and 4-D, and decodes 4-5× faster. TTHRESH, the only baseline that runs natively in 5-D and 6-D, encodes 5-7× slower than LATEC there.
LATEC is therefore practical for the many workloads, in particular when data is compressed once and rarely read. Simulation checkpoints are a typical example: they are decompressed only if the simulation must be restarted, so rate matters much more than decoding time.
W2, Q1. Domain gap, fine-tuning, and per-input training.
Per-input training helps only after many steps, and its training time is far larger than the compression time; after 50-200 steps, the stream is up to 20% larger. Fine-tuning once per dataset gives comparable or larger gains at no cost per input. This experiment is now in App. C.4.
Q3. Gating versus the GNN alone (new Figs. 6 and 8, App. E.2). We thank the reviewer for suggesting this ablation. We compare the GNN alone with LATEC (GNN + gate) on crops at RME 1e-3. Values are % bpp vs SZ3; lower is better:
Q4. Model capacity (new App. E.2.4, Table E.3). To evaluate the impact of model capacity, we retrained the model at widths 16 and 32 with the full recipe and ran the full benchmark, excluding Gysela. Speed-ups are relative to width 64, on the H100 setup of Fig. 9. The last column is the mean % bpp vs SZ3 over these 12 fields (width 64 averages −18.1% over all 13 fields of Table 2); per-field values are in Table E.3. Lower is better:
Width 32 keeps most of the gain (4.5% more bits than width 64), but it is only 1.0-1.6× faster to encode and 1.3-1.5× faster to decode. Runtime is dominated by per-point work that does not depend on the width, such as building each point's causal context. Width 16 is far too small: it uses more bits than SZ3 on four fields, and 11× as many bits on S3D. Width 32 remains a useful trade-off when decoding speed matters most, and we will release it alongside the width-64 model in the LATEC code.