{"id":"96a73913-2169-4f2d-8e52-264663c7b140","arxiv_id":"2502.02593","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A diffusion transformer with plane position embeddings reconstructs full 3D velocity fields from exact 2D slices of DNS flow data, with the largest accuracy gains on an easy interpolation benchmark.","lead":"This paper trains a diffusion transformer to reconstruct a 3D flow field from one or two 2D slices of that flow. The authors claim this could lower the cost of particle-image-velocimetry experiments and speed up flow analysis.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The flagship INS(INT) accuracy (nRMSE 0.0053) uses a random split of DNS time steps with Δt=0.07, so near-duplicate training snapshots may make the result retrieval-like; the honest extrapolation split gives only nRMSE 0.133.","rationale":"I read the paper's central claim as: a diffusion transformer conditioned on two orthogonal DNS slices can recover the 3D velocity field, with near-perfect accuracy. For that to be true, the test must measure generalization to unseen flow fields, not retrieval of near-duplicate training snapshots. The INS(INT) protocol does not guarantee this because random splitting with Δt=0.07 leaves test snapshots adjacent to training snapshots; the paper even admits consecutive snapshots are similar. This is the weakest load-bearing point because every strong quantitative statement in the abstract and conclusion points to INS(INT), while the extrapolation results (INS(EXT), CNS) show only modest gains over F-FNO. The attention ablation's Small-model degradation and the absence of code or error bars are additional concerns, but they are secondary; even if the attention claim were softened, the reconstruction claim would still need a temporally clean evaluation. I keep a CONDITIONAL verdict with the added condition that the authors rerun INS with a proper temporal gap and report autocorrelation-based separation; if the result collapses, the current claims should be substantially weakened.","tokens_in":16256,"tokens_out":7957,"duration_ms":77677,"concrete_test":"Re-evaluate Base (and Large*) on the INS dataset using a strict temporal hold-out: train on time steps up to T0, test on time steps greater than T0, and discard a buffer of at least one large-eddy turnover time between train and test, with the buffer size justified by the temporal autocorrelation of the velocity field. Report the minimum train-test time separation and the autocorrelation values. If nRMSE remains below 0.02 on this split, the interpolation result is genuine; if it rises to the ~0.13 level seen in INS(EXT), the headline accuracy claim should be restricted to interpolation or removed.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The most load-bearing weakness is the evaluation protocol behind the headline number, not the architecture. In Table 2, the Base model reaches nRMSE 0.0053, PSNR 51.0158, and SSIM 0.9997 on INS(INT), but Section 4.1 defines INS(INT) as a random split of DNS snapshots, with consecutive snapshots only Δt=0.07 apart. In a turbulent channel flow at Reτ=180, 0.07 is far below the correlation time, so a randomly held-out test snapshot will typically have a training snapshot a few Δt away that is nearly identical. The network can then satisfy the two-plane conditioning by retrieving that near-duplicate volume rather than by learning a general mapping from 2D slices to 3D structure. The paper's own INS(EXT) split removes this overlap by training on the first 80% of the time series, and there the Base model drops to nRMSE 0.1331, essentially tied with F-FNO (0.1421). The abstract's unconditional statement that the model 'can efficiently and accurately reconstruct 3D flow fields from 2D flow data' is therefore carried almost entirely by an interpolation benchmark whose random split permits temporal near-duplication. The central claim would still be interesting if it held on extrapolation, but the reported extrapolation numbers are only marginally better than a standard neural operator baseline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a diffusion-transformer method for reconstructing a full 3D velocity field from a small set of 2D planar slices. The model is conditioned in three ways: by padding and channel-concatenating the input slices with the volume being denoised, by injecting CLIP image features and plane position embeddings through adaptive layer norm, and by cross-attending to the CLIP encoder output. Global attention is replaced with window attention and plane attention to reduce the cost of processing 3D voxels. Experiments are carried out on DNS data for a turbulent channel flow at Re_tau=180 (INS) and compressible Navier-Stokes data (CNS), plus a geometry dataset, and the model is compared with F-FNO, U-Net-mod, Dil-ResNet, and 2D3DGAN. On the random-split interpolation benchmark INS(INT), the Base model reports nRMSE 0.0053, PSNR 51.0158, and SSIM 0.9997; on the temporally extrapolative split INS(EXT), the same model reports nRMSE 0.1331 versus 0.1421 for F-FNO. The paper concludes that the model can efficiently and accurately reconstruct 3D flow fields from 2D data.","tokens_in":16423,"tokens_out":4648,"duration_ms":44423,"significance":"If the central claim were fully supported, the method would be practically useful for reducing the cost of three-dimensional PIV and CFD studies, and the plane position embedding plus window/plane attention are reasonable architectural contributions. The paper is honest about some limitations, such as the sensitivity to classifier-free guidance, and it provides several ablations, including attention type and patch size. However, the headline accuracy is achieved on a benchmark that the paper itself acknowledges contains near-duplicate training and test snapshots, and on the temporally separated benchmark the gain over F-FNO is marginal. The paper does not ship code, data, or error bars, and diffusion sampling is stochastic, so the reported single point estimates are not enough to establish the comparative claims. The significance is therefore conditional: the architecture is plausible, but the evidence for accurate 3D reconstruction from 2D data is not yet convincing.","major_comments":[{"comment":"The INS(INT) benchmark randomly splits DNS snapshots that are only Delta t = 0.07 apart. For a turbulent channel flow at Re_tau=180, this interval is far below the correlation time, so each held-out test snapshot almost certainly has a near-duplicate in the training set. The paper itself states that 'the flow fields at consecutive time steps are similar' (Section 4.1). Because the exact 2D slices are supplied as conditioning, the model can succeed on this split by retrieving a nearly identical training volume rather than by learning a general 2D-to-3D mapping. Consequently the headline numbers nRMSE 0.0053, PSNR 51.0158, SSIM 0.9997 do not measure generalization, and the abstract's unconditional claim that the model 'can efficiently and accurately reconstruct 3D flow fields from 2D data' rests on this contaminated benchmark. The temporally separated INS(EXT) split is the relevant generalization test, and there the Base model achieves only nRMSE 0.1331 versus 0.1421 for F-FNO, which is a marginal improvement. The paper should either report the minimum temporal gap between training and test snapshots in the interpolation split or replace it with a split that guarantees separation, and the conclusions should be scaled to the extrapolation results.","section":"Section 4.1, Table 2"},{"comment":"All reported metrics are single point estimates with no error bars, no multiple seeds, and no statement of the number of diffusion sampling runs averaged. Since the method is a stochastic generative model, the differences between models on the extrapolation benchmarks, which are often only 0.005-0.01 in nRMSE, may be within sampling noise. The paper should report means and standard deviations over multiple training seeds and multiple sampling runs, and should state whether the differences against F-FNO are statistically significant.","section":"Section 4.3, Tables 2-6"},{"comment":"The conditioning inputs are exact, noiseless, same-grid DNS slices, while the abstract and introduction motivate the method by PIV experiments, which produce noisy, lower-resolution, and potentially misaligned measurements. The paper does not test robustness to noise, downsampling, or misalignment, so the practical claim that the method can reduce PIV cost is not supported by the experiments. At minimum, a sensitivity study with synthetic noise or degraded slices is needed, or the claims should be restricted to noiseless simulation data.","section":"Section 3.3, Section 4.1"},{"comment":"The plane position embedding is one of the three listed contributions, but the ablation in Figure 7 does not isolate its effect: the model is evaluated with three planes at a fixed relative position of -5, and the metrics are plotted against spatial position, but there is no comparison against a model without position embedding, nor against a fixed one-hot or integer plane index. The observed high quality near the reference planes could be driven by the supplied slice content rather than by the learned position encoding. The paper should include a with-versus-without comparison for the plane position embedding across multiple plane configurations.","section":"Section 4.5, Figure 7"}],"minor_comments":[{"comment":"\"Here, We propose\" has an erroneous capital W; please fix the capitalization and also re-read the abstract for other grammatical slips such as \"computationalresources\". ","section":"Abstract"},{"comment":"The norm notation in Equation (2) is malformed: it appears as \"∥...∥2^2\" without a clear subscript, and the expectation is missing the appropriate parentheses. Please rewrite it as \\(\\mathbb{E}_{t,S_0,\\epsilon}\\left[\\|\\epsilon - \\epsilon_\\theta(S_t, c(t,P,E_P))\\|_2^2\\right]\\) or an equivalent form.","section":"Equation (2)"},{"comment":"In the CNS column, the Large* model reports PSNR 25.00220 with an extra decimal place; please make the precision consistent. Also, the text in Section 4.3 says \"our Large model may exhibit overfitting,\" but Table 2 reports results for \"Large* (Ours)\"; clarify whether the overfitting comment refers to the global-attention Large or the window/plane-attention Large*.","section":"Table 2"},{"comment":"The column heading \"Training Relative Promotion (%)\" is unclear; replace it with a concrete measure such as training wall-clock time or steps per second, and report absolute values with variance. The phrase \"promotion\" is not standard in this context.","section":"Table 5"},{"comment":"The caption \"CNSxOy planes where the reference xOy plane is not provided\" is confusing; please rephrase to make clear which planes are inputs and which are outputs for each panel.","section":"Figure 3"},{"comment":"The F-FNO configuration is described as a 24-layer network with 16 Fourier modes and 64 channels, but no citation or appendix section is given for this exact setup. Please provide a reference or a detailed configuration table.","section":"Section 4.2"},{"comment":"\"Pacth Size\" is a typo for \"Patch Size.\" Also, when reporting that the model fails to converge for patch size 8, state the training budget and the criterion used to declare non-convergence.","section":"Table 6"}],"recommendation":"major_revision","confidential_remarks":"This is a technically plausible but under-validated submission. The strongest reported result is driven by a random split that the authors themselves admit contains near-duplicate snapshots; the extrapolation results are only marginally better than a standard FNO baseline, and no error bars are provided despite stochastic sampling. I recommend major revision rather than rejection because the architecture and ablations could support a defensible paper if the evaluation is redone with temporally separated splits, repeated seeds, and robustness tests. The authors do not provide code or data, which further limits reproducibility; encouraging a public release would strengthen the contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the genuinely new piece here is the plane position embedding, which lets the model condition on 2D slices at arbitrary positions, not just fixed planes like the prior work [41,76]. The diffusion transformer with window and plane attention is a sensible adaptation, and the paper includes useful ablations on number of planes, patch size, and attention, plus a geometry test. The writing is clear, and the authors are upfront that INS(INT) is the \"ample data\" split.\n\nThe soft spots are load-bearing. The flagship INS(INT) number (Base nRMSE 0.0053) is almost certainly a near-duplicate retrieval effect: with Δt=0.07 in a Reτ=180 channel, consecutive DNS snapshots are well within the correlation time, so a randomly held-out test snapshot has a near-twin in the training set. The paper's own Section 4.1 says exactly this. On the honest extrapolation split, Base nRMSE is 0.1331 versus F-FNO's 0.1421—a marginal edge. The abstract's unconditional \"accurately reconstruct\" claim is carried almost entirely by the interpolation benchmark.\n\nThe attention ablation also deserves more scrutiny. For the Small model, replacing global attention degrades nRMSE from 0.1298 to 0.1389 and SSIM from 0.8132 to 0.7887. Calling that \"minimal impact\" is generous. The Large model improves, so the story is mixed, and the paper does not discuss why.\n\nOther concerns: all metrics are single point estimates with no error bars or multiple seeds, which matters because diffusion sampling is stochastic. No code or weights are provided, so the central numbers are not independently verifiable. The model is trained on clean, perfectly aligned DNS slices; real PIV data will be noisy, downsampled, and misregistered, and the paper tests none of that. The CLIP conditioning is a reasonable choice, but it is not ablated, so its contribution is unknown.\n\nBottom line: the core idea is worth pursuing, and the paper deserves a serious referee, but it is not ready to be accepted as is. I would send it out with a request for major revision: release code and weights, add error bars and multiple seeds, test on noisy or downsampled inputs, compare against a simple inpainting baseline, and temper the claims to match the extrapolation results. A fixed version would be useful to people working on flow reconstruction from partial measurements; the current version oversells what the evidence shows.","headline":"A plausible engineering contribution whose headline accuracy number is inflated by the interpolation split; the honest extrapolation result is only marginal, so the paper needs major revision before its claims can be trusted.","tokens_in":17094,"tokens_out":1777,"would_cite":false,"duration_ms":16297,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A diffusion transformer conditioned on two orthogonal 2D velocity slices can reconstruct the full 3D turbulent flow field, reaching near-simulation accuracy on in-distribution DNS data.","keywords":["3D flow reconstruction","diffusion transformer","plane position embedding","particle image velocimetry","turbulent channel flow","window attention","plane attention","voxel space"],"falsifier":"A concrete experiment: take the INS(INT) test set, add Gaussian noise with standard deviation of 1-5% of the mean velocity to the two input slices, and measure nRMSE; if the error rises sharply from the clean-slice value of 0.0053, the noiseless-slice assumption is load-bearing.","tokens_in":15872,"feed_emoji":"🌊","tokens_out":8921,"duration_ms":71983,"temperature":0.7,"pith_summary":"This paper tries to establish that a diffusion transformer can recover a full three-dimensional velocity field from just two orthogonal two-dimensional slices of it. If true, experimentalists could get 3D flow data from cheaper 2D particle-image velocimetry instead of 3D setups with multiple cameras, and CFD users could reduce reliance on expensive full-volume simulations. The authors train the model on direct numerical simulation data, condition it on the two slices through several pathways, and add plane position embeddings so the same model can handle slices at any location. On the in-distribution interpolation benchmark their Base model reaches a normalized RMSE of 0.0053 and an SSIM of 0.9997, and the model still produces meaningful flow when one reference plane is withheld.","feed_headline":"Two 2D flow slices rebuild the full 3D field at 0.0053 error","feed_subtitle":"A diffusion transformer fills the unmeasured volume from any slice pair, potentially cutting 3D flow-measurement cost.","key_machinery":"The load-bearing identity is the simplified denoising objective $$L_{\\text{simple}}(\\$\\theta$)=\\mathbb{E}_{t,S_0,\\epsilon}\\|\\epsilon-\\epsilon_\\$\\theta$(S_t,c(t,P,E_P))\\|$_2^{2}$,$$ which trains the transformer to predict the noise added to the 3D volume while conditioned on the 2D planes $P$ and their positions $E_P$. Two mechanisms make this practical: plane position embeddings, which are Fourier features of the normalized plane coefficients $A',B',C',D'$ in the plane equation $A'x+B'y+C'z+D'=0$, and two linear-complexity attention schemes. Window attention groups tokens into $w\\times w\\times w$ cubes, reducing complexity from $O(L^2)$ to $O(Lw^3)$, and plane attention lets tokens attend only within the $yOz$, $xOz$, and $xOy$ planes. Together they give the model slice-location awareness while keeping 3D token processing affordable.","core_discovery":"The central claim, stated in the authors' terms, is that a voxel-space diffusion transformer can act as a learned prior for turbulent flows: it fills in the unmeasured volume between and beyond two input slices with near-ground-truth accuracy when the flow regime is covered by training data. The conditioning has three streams: the 2D slices are padded and channel-concatenated with the noisy 3D volume, a pretrained image encoder's global and final-layer features are injected through adaptive layer norm and cross-attention, and the normalized plane equation coefficients are Fourier-embedded so the model knows each slice's position. The plane position embedding is what enables reconstruction from arbitrary slice combinations rather than a fixed measurement geometry.","pith_inferences":["Inference: because the conditioning slices are exact voxels, the 0.0053 nRMSE result is likely an upper bound on real-world performance; adding PIV-like noise or slight slice misalignment would degrade it, and the size of that degradation is the real test of practical value.","Inference: the conditioning uses a pretrained image encoder whose representations were learned on photographs, which is a domain mismatch for multi-channel velocity slices; a physics-specific encoder or one based on local flow statistics might be simpler and equally effective.","Inference: the plane position embedding parameterizes any plane intersecting the unit cube, so the same framework could extend to non-axis-aligned cutting planes, non-uniform sensor placements, or time-resolved reconstruction.","Inference: the quality-versus-distance results imply an optimal sensor-placement strategy: multiple evenly spaced reference planes give better coverage than many planes clustered near one location, which could directly inform PIV experiment design."],"forward_implications":["On the in-distribution interpolation benchmark, the Base model reaches nRMSE 0.0053, PSNR 51.0158, and SSIM 0.9997, meaning the reconstructed 3D volume is numerically and visually nearly indistinguishable from the DNS ground truth.","A single trained model can reconstruct from arbitrary plane combinations because plane position embeddings encode where each slice lies; adding an extra plane does not always improve quality and in some cases slightly hurts it.","Window and plane attention speed up training by 27.9-39.3% across model sizes while barely changing accuracy, and for the Large model the efficient attention variant actually outperforms global attention.","On extrapolation benchmarks (INS(EXT), CNS), the method remains competitive and its SSIM lead indicates better-preserved small-scale structure, though reconstruction quality degrades with distance from the input planes.","When one reference plane is withheld, the diffusion model still produces a meaningful flow field, whereas the strongest non-diffusion baseline produces meaningless noise on that plane."],"supporting_citations":[{"why":"Supplies the incompressible turbulent channel-flow dataset (friction Reynolds number 180) and the GAN-based baseline that the method must beat.","marker":"[76]"},{"why":"Supplies the compressible Navier-Stokes dataset used for the extrapolation test.","marker":"[64]"},{"why":"Provides the diffusion-transformer backbone design, including patchify, adaptive layer norm, and zero-initialized MLPs, which the model adapts to 3D flows.","marker":"[51]"},{"why":"Defines the DDPM training objective and iterative denoising process that the reconstruction is built on.","marker":"[23]"},{"why":"Supplies the Transformer architecture and attention formulation, including the cross-attention layer.","marker":"[70]"},{"why":"Shows how to adapt diffusion transformers to 3D voxel data with 3D patchify and 3D positional embeddings.","marker":"[45]"},{"why":"Provides the 2x2-large geometry dataset used to test reconstruction around obstacles.","marker":"[35]"},{"why":"F-FNO is the strongest non-diffusion baseline; the paper compares against it on all three benchmarks.","marker":"[67]"}],"fun_headline_variants":["Diffusion transformer maps 2D slices to full 3D flow","3D flow from any two 2D slices, error 0.0053","Learned prior fills unmeasured volume from 2D slices","From 2D PIV to full 3D flow: one transformer","Slice pair to 3D: diffusion transformer achieves 0.0053"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the two input planes are exact, noiseless voxel slices of the target 3D field; if real PIV data are noisy, downsampled, or misaligned, the model has not been shown to handle it.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion transformer maps 2D slices to full 3D flow","3D flow from any two 2D slices, error 0.0053","Learned prior fills unmeasured volume from 2D slices","From 2D PIV to full 3D flow: one transformer","Slice pair to 3D: diffusion transformer achieves 0.0053"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000207,"raw_usage":{"total_tokens":1365,"prompt_tokens":878,"completion_tokens":487,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":388}},"tokens_in":494,"tokens_out":487,"duration_ms":3883,"temperature":1.0,"reasoning_tokens":388,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:00:56.913141+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete experiment: take the INS(INT) test set, add Gaussian noise with standard deviation of 1-5% of the mean velocity to the two input slices, and measure nRMSE; if the error rises sharply from the clean-slice value of 0.0053, the noiseless-slice assumption is load-bearing.","supporting_citations":[{"cited_title":"From zero to turbulence: Generative modeling for 3d flow simulation","cited_arxiv_id":null,"evidence_quote":"Provides the 2x2-large geometry dataset used to test reconstruction around obstacles."},{"cited_title":"A deep-learning approach for reconstructing 3d turbulent flows from 2d observation data","cited_arxiv_id":null,"evidence_quote":"Supplies the incompressible turbulent channel-flow dataset (friction Reynolds number 180) and the GAN-based baseline that the method must beat."},{"cited_title":"Pdebench: An exten- sive benchmark for scientific machine learning.Advances in Neural Information Processing Systems, 35:1596–1611, 2022","cited_arxiv_id":null,"evidence_quote":"Supplies the compressible Navier-Stokes dataset used for the extrapolation test."},{"cited_title":"Dit-3d: Exploring plain diffusion transformers for 3d shape generation","cited_arxiv_id":null,"evidence_quote":"Shows how to adapt diffusion transformers to 3D voxel data with 3D patchify and 3D positional embeddings."},{"cited_title":"Factorized fourier neural operators","cited_arxiv_id":null,"evidence_quote":"F-FNO is the strongest non-diffusion baseline; the paper compares against it on all three benchmarks."}],"review_version":1}