{"id":"1fe82adc-67d6-4494-893a-b6c30ad5206b","arxiv_id":"2508.04318","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A Reynolds-conditioned diffusion model can generate DHIT turbulence boxes for LES/DNS inflow that match energy spectra and development length, though integral length scale and anisotropy are imperfect.","lead":"This paper trains a diffusion model to generate 3D turbulence boxes and injects them at the inlet of scale-resolving simulations, offering a memory-saving alternative to storing large precursor libraries. It finds the generated fluctuations reproduce energy spectra and development distance in free-domain tests, while integral length scale and Reynolds-stress anisotropy deviate.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Scalar Re_Lint conditioning does not encode L_int: generated samples show ~14–20% L_int error and nearly identical correlation functions across Re_Lint, undermining the claimed reproduction of two-point statistics.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing premise: a single scalar Re_Lint is asked to determine the turbulence statistics, in particular the integral length scale. My reading of the full text strengthens this concern with direct empirical evidence from the authors: the a priori section admits near-identical correlation functions at different Re_Lint, and the a posteriori figures show 14-20% L_int errors. The strongest claim about energy-spectrum overlap after less than one box length is not directly invalidated, and the paper deserves credit for the a posteriori free-domain injection test and the MMPS continuity construction. However, the broader central claim—that generated samples accurately reproduce turbulence statistics and enable easy parametrization of inlet turbulence—is compromised if L_int cannot be independently controlled or reproduced. This is not an internal inconsistency in the diffusion model, but a limitation of the conditioning variable. Since the reader already returned a CONDITIONAL verdict, my concern does not change the verdict; it sharpens the condition that must be met: either demonstrate that L_int is a deterministic function of Re_Lint in the training distribution, or condition on (k, L_int) jointly and re-validate.","tokens_in":26888,"tokens_out":6093,"duration_ms":74963,"concrete_test":"On the 214-box database, plot L_int against Re_Lint (Eq. 11) and measure the residual spread around a fit L_int = f(Re_Lint). If boxes with similar Re_Lint differ in L_int by more than ~10%, the scalar condition is non-injective. Then generate 50 samples at each of 10 evenly spaced Re_Lint values in the training range, compute L_int from the generated autocorrelation functions via Eq. 3, and compare each generated distribution with the target L_int. If the generated L_int is approximately constant across Re_Lint, or if the relative error exceeds 10% at most target values, the conditioning premise fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is Section 4's conditioning on the scalar Re_Lint = sqrt(2/3 k) L_int / nu (Eq. 11), which the paper asserts embeds both TKE and integral length scale. The paper's own evidence contradicts this. Section 4.1 states: 'the model seems to have difficulty distinguishing between the integral length scales and predicts the same correlations at each ReLint.' Section 4.2.1 reports L_int overestimated by about 14%; Section 4.2.2 reports 'almost the same Lint at each TKE level (about 20% of relative error at x/L = 4).' Thus, even within the training range, the generated samples do not reproduce the target length scale, and the abstract's claim that two-point autocorrelation functions are accurately reproduced is not supported. If L_int is not a deterministic function of Re_Lint in the training data, or if the model cannot vary L_int at fixed Re, a user specifying Re cannot target a desired vortex size. This is an identifiability/conditioning insufficiency, not a failure of diffusion sampling per se. A leftover caption note in Figure 9 ('Discuss more the limitation of a pure data-driven model...') further indicates the issue is acknowledged but not resolved.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a classifier-free diffusion model trained on 214 three-dimensional snapshots of Decaying Homogeneous Isotropic Turbulence to generate instantaneous turbulent velocity boxes for use as inflow boundary conditions in LES/DNS. The model is conditioned on a scalar Reynolds number Re_Lint = sqrt(2/3 k)L_int/nu. Generated boxes are assessed a priori via energy spectra, two-point autocorrelations, barycentric anisotropy triangles, and vorticity PDFs, and a posteriori by injecting them into a free domain with two concatenation methods (blending and moment matching posterior sampling). The authors compare against a library-based precursor method and report similar development distances, lower storage, and faster turnaround for finding target TKE levels.","tokens_in":27296,"tokens_out":6324,"duration_ms":62927,"significance":"The manuscript addresses a real bottleneck in scale-resolving simulations: the high memory and setup cost of precursor-based turbulence inflow. It demonstrates that a diffusion model can generate 3D turbulent boxes with reasonable energy spectra and TKE at multiple Reynolds numbers, and that injected samples develop qualitatively as quickly as the precursor method. The computational cost comparison (Table 2, Fig. 21) and the use of MMPS for temporal continuity are useful contributions. However, the central claim about accurate reproduction of two-point statistics is not supported by the reported results, and the conditioning on a single scalar is not validated for length-scale control. With revisions, the method remains a promising proof-of-concept.","major_comments":[{"comment":"The abstract claims samples 'accurately reproduce turbulence statistics, such as the energy spectrum and the two-point autocorrelation functions.' The paper's own a priori results contradict this: §4.1 states there is 'an underestimation of f(r) and an overestimation of g(r)' and that 'The model seems to have difficulty distinguishing between the integral length scales and predicts the same correlations at each ReLint.' A posteriori, §4.2.1 reports L_int overestimated by about 14%; §4.2.2 reports about 20% relative error at x/L=4. Please quantify the errors on f(r), g(r), and L_int and revise the abstract and conclusions accordingly. This is not a presentation issue; it is the paper's headline result.","section":"Section 4.1 / Fig. 10 / Abstract"},{"comment":"The conditioning strategy assumes a single scalar Re_Lint = sqrt(2/3 k) L_int / nu uniquely determines the target statistics. The paper does not show identifiability: in the 214 training boxes, is L_int a deterministic function of Re_Lint? If two different (k, L_int) states share a Re_Lint value, or if the model cannot vary L_int at fixed Re_Lint, then a user specifying Re_Lint cannot target a desired vortex size. The failure to reproduce L_int (Figs. 10, 14, 18) is consistent with this identifiability problem. Please plot L_int vs Re_Lint for the training data and consider conditioning on TKE and L_int separately (e.g., two-channel conditioning) or justify why Re_Lint is sufficient.","section":"Section 4, Eq. (11)"},{"comment":"The hyperparameters lambda*, tau*, gamma* and the MMPS parameters m, Sigma_y are selected by minimizing a loss (Eq. 12) that includes the same statistics (k, E(kappa), f/g, barycentric coordinate) used later for validation. Consequently, the reported a priori agreement is not independent evidence of generalization. Please perform a nested validation (e.g., a separate hyperparameter-selection split) or report sensitivity of the metrics to lambda, tau, gamma, m, Sigma_y. Note that the h_{f,g} term is weighted by 100, so the persistent failure of correlation functions is especially telling.","section":"Section 4.1, Eq. (12) and §3.2"},{"comment":"The a posteriori comparison rests on only two test conditions and on single trajectories without error bars or statistical significance over the 24 generated boxes. The statement 'The development distance of the two methods is approximately the same' (Section 5) is qualitative. Please define the development length quantitatively (e.g., distance to reach within a specified tolerance of the target TKE and spectrum) and report uncertainties. This is the main practical advantage claimed for the method.","section":"Sections 4.2.1-4.2.2 / Section 5"}],"minor_comments":[{"comment":"The caption contains a leftover internal note: 'Discuss more the limitation of a pure data-driven model...' Remove it before submission.","section":"Figure 9 caption"},{"comment":"The definition of R_ij appears with a comma inside the ensemble average; it should be R_ij(r,t) = <u_i(x+r,t) u_j(x,t)>.","section":"Eq. (5)"},{"comment":"The text counts 214 boxes and then says 62 boxes and 17 boxes are used for training and validation; the relationship to Table 1's training size 136,102 (after translations) should be stated clearly.","section":"Section 2.3.1 / Table 1"},{"comment":"Both captions say 'three streamwise locations,' but the figures show eight panels. Correct the captions to match the displayed x/L values.","section":"Figure 15 / Figure 20 captions"},{"comment":"The symbols ✓, ≈, and ✗ need a legend; the Storage row formatting (32MB GPUs, 72MB CPUs, 3MB CPUs) is confusing and should be clarified.","section":"Table 2"},{"comment":"'descent synthetic turbulence' should be 'decent.' The conclusion should be toned down to reflect the L_int limitation rather than stating 'good agreement' without qualification.","section":"Section 2.3.2 / Section 6"}],"recommendation":"major_revision","confidential_remarks":"The abstract states the article has been 'submitted to/accepted by Physics of Fluid,' which is unusual in a preprint under review and may confuse readers. The leftover caption note in Figure 9 also suggests the manuscript was not fully cleaned. I recommend requesting removal. The main issue is an overclaim that can be fixed via revision: the method may still be valuable for TKE/spectrum control, but the length-scale limitation must be acknowledged and the conditioning scheme re-examined."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a useful proof-of-concept, but the headline claim is at odds with the paper's own evidence. What is new: the authors adapt classifier-free diffusion to generate 3D DHIT boxes conditioned on a single Reynolds number, use MMPS to make consecutive boxes continuous, and show in a real DG-LES solver that the synthetic inlets develop as fast as precursor-based inlets while avoiding stored precursor libraries. That practical target is real, and the a posteriori spectrum overlap within one box length is credible.\n\nCredit where it's due: they compare against a library-based precursor in a production solver, include two a posteriori cases, report training and inference costs, and cite the relevant diffusion-inflow work. The MMPS time-correlation test is a sensible check. The scope is honest about being DHIT-only, and the paper does not pretend to solve wall-bounded or anisotropic cases.\n\nSoft spots, in increasing order. First, the abstract says samples accurately reproduce two-point autocorrelation functions; Section 4.1 says the opposite: f(r) is underestimated, g(r) is overestimated, and the model predicts nearly the same correlations at each ReLint. The a posteriori integral-scale errors are about 14% and 20%. That is a direct internal contradiction. Second, the conditioning variable is the real issue. ReLint = sqrt(2/3 k) Lint / nu is a single scalar, and the generated samples do not vary Lint in the way the target states do. If a user wants a specific vortex size, this model cannot yet deliver it. The leftover note in the Figure 9 caption reinforces that the authors know this is unresolved. Third, the optimizer hyperparameters and MMPS parameters are chosen by minimizing loss built from the same statistics used for validation, and Barycentric outliers are simply discarded; both inflate the apparent a priori quality. Missing error bars make the quantitative claims hard to interpret. These are real but fixable: use more independent metrics, report variability, and soften the claims.\n\nWho this is for: people working on ML-generated inflow conditions, especially those combining generative models with CFD. It deserves a serious referee rather than a desk reject. If I were editing, I would send it out and ask for claims to be aligned with evidence before acceptance. As it stands, it is a promising engineering demonstration, not a validated method for controlling turbulent length scales via ReLint.","headline":"A promising ML-CFD proof-of-concept whose storage-free inflow generation works reasonably well, but the abstract overstates autocorrelation accuracy and the scalar Re_Lint conditioning does not actually control L_int.","tokens_in":27752,"tokens_out":2894,"would_cite":true,"duration_ms":37673,"reading_group":"maybe","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 model conditioned on one Reynolds number can generate realistic turbulent inlet fields for LES/DNS, replacing stored precursor libraries.","keywords":["turbulent inflow generation","diffusion models","decaying homogeneous isotropic turbulence","large eddy simulation","classifier-free guidance","inlet boundary condition","moment matching posterior sampling","energy spectrum"],"falsifier":"Generate two DHIT boxes with the same $Re_{L_{\\mathrm{int}}}$ but different pairs $(k, L_{\\mathrm{int}})$ — for example, double $k$ and halve $L_{\\mathrm{int}}$ so the product is unchanged — and condition the model at that Reynolds number. If the model produces the same energy spectrum and integral length scale for both, the conditioning is degenerate and the central premise fails.","tokens_in":26823,"feed_emoji":"🌪️","tokens_out":7276,"duration_ms":83405,"temperature":0.7,"pith_summary":"Turbulent inflow remains a bottleneck for scale-resolving simulations: precursor/library methods give realistic fluctuations but require expensive auxiliary simulations and large stored libraries, while synthetic generators are cheap but need a development distance. This paper tries to get the realism of the precursor method without its memory and turnaround cost by training a classifier-free diffusion model on decaying homogeneous isotropic turbulence (DHIT) boxes. The model is conditioned on a single scalar, $Re_{L_{\\mathrm{int}}} = \\sqrt{2/3}\\,k\\,L_{\\mathrm{int}}/\\nu$, which encodes both the turbulent kinetic energy level and the integral length scale. On injection into a free domain, the energy spectrum of the synthetic boxes overlaps that of the true DHIT within about one box length, matching the development distance of the precursor method while storing only the model (about 32 MB) instead of 24 flow fields. The paper is explicit that the hardest statistic to reproduce is the two-point autocorrelation function, with the generated integral length scale overestimated by about 14%.","feed_headline":"Diffusion model generates inlet turbulence without stored libraries","feed_subtitle":"Synthetic boxes match precursor energy spectra within one box length while cutting storage to a 32 MB model.","key_machinery":"The carrying mechanism is a classifier-free guided diffusion model: a UNet score network trained with the variance-preserving SDE objective to denoise $3\\times64^3$ velocity boxes, conditioned on the scalar Reynolds number $Re_{L_{\\mathrm{int}}} = \\sqrt{2/3}\\,k\\,L_{\\mathrm{int}}/\\nu$. This single number is the interface between the user-specified turbulence state and the generated field — it is what lets the model interpolate across TKE/scale combinations without retraining. Temporal continuity between successive generated boxes, required for an inlet condition, is imposed by Moment Matching Posterior Sampling, which treats the last $m=16$ slices of the previous box as a noisy measurement co","core_discovery":"A classifier-free guided diffusion model trained on 214 DHIT boxes spanning $Re_{L_{\\mathrm{int}}} \\in [196.58, 1302.16]$ can generate new $64^3$ instantaneous velocity boxes whose a priori statistics (energy spectrum, vorticity distribution, Reynolds-stress anisotropy) closely match ground truth, and after injecting them at the inlet of a free-domain LES the turbulence develops as fast as with the original precursor method — in less than one box length for the energy spectrum at 23.3% turbulence intensity. The conditioning variable $Re_{L_{\\mathrm{int}}}$ embeds both the turbulent kinetic energy level and the integral length scale, so a single model covers many target states and interpolate","pith_inferences":["Not tested here: because conditioning is a single scalar, any two physical states with the same $Re_{L_{\\mathrm{int}}}$ but different $(k, L_{\\mathrm{int}})$ are indistinguishable to the model; conditioning separately on $k$ and $L_{\\mathrm{int}}$ would directly test whether this embedding is lossy.","The observed asymmetry — extrapolating to lower $Re_{L_{\\mathrm{int}}}$ works well, extrapolating to higher does not — suggests a learned hierarchy in which lower-Reynolds fields are filtered versions of higher-Reynolds ones. A testable consequence is that a model trained on a longer decay sequence could generate new states by denoising downward rather than by inventing small scales.","The MMPS inlet anisotropy (dominant $u'u'$ component) indicates that conditioning on $y$–$z$ slices constrains the streamwise velocity differently from the cross-stream components; adding a divergence-free projection during sampling, which the paper lists as a perspective, would be a direct test.","Because the training set is built from periodic boxes with translation augmentation, the current proof is tied to homogeneous freestream turbulence; extending to boundary layers or wakes would require a non-periodic architecture and a different training representation."],"forward_implications":["Within the trained Reynolds-number range, the user can query any target TKE/length-scale state without running a new DHIT; 24 boxes are generated in about three minutes on the tested GPU setup.","The inlet library no longer needs to be stored or read from disk: the model itself is the library, and successive boxes can be generated on demand.","The development distance of the generated turbulence is not increased relative to the precursor method (1–2 box lengths, depending on turbulence intensity), so no extra fetch length must be added to the computational domain.","The stochasticity of the diffusion model decorrelates successive inlet boxes: MMPS keeps the time correlation below 10%, unlike simply re-injecting one frozen DHIT box.","The same model can be reused in the trial-and-error loop that matches an inlet condition to measurements at a target station, because regenerating at a new conditioning value costs seconds instead of a full precursor rerun."],"supporting_citations":[{"why":"Defines denoising diffusion probabilistic models, the generative backbone used to synthesize the turbulence boxes.","marker":"Ho et al. (2020)"},{"why":"Supplies the predictor/corrector sampling scheme and the Moment Matching Posterior Sampling procedure used for temporal continuity between boxes.","marker":"Rozet and Louppe (2023)"},{"why":"Describes the precursor/library boundary-condition method with translations, rotations and blending that the diffusion approach is designed to replace.","marker":"Rasquin et al. (2023)"},{"why":"Sets the evaluation criteria (realism, development length, storage, computational cost) against which the method is compared.","marker":"Dhamankar et al. (2015)"},{"why":"Provides classifier-free guidance, the conditioning mechanism that lets one model generate fields at different Reynolds numbers.","marker":"Ho and Salimans (2022)"},{"why":"Supplies the blending procedure used to concatenate consecutive boxes while conserving turbulent kinetic energy.","marker":"Xiong et al. (2012)"},{"why":"Frames diffusion in continuous time as an SDE and derives the probability-flow ODE underlying the fast sampling used here.","marker":"Song et al. (2021a)"}],"fun_headline_variants":["Diffusion model cuts inlet turbulence storage to 32 MB","AI generates turbulent inflow without bulky libraries","One diffusion model replaces entire turbulence database","Synthesize inlet turbulence on the fly with diffusion","32 MB model churns out realistic turbulence for CFD"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that a single Reynolds number $Re_{L_{\\mathrm{int}}}$ contains enough information to fix the turbulence statistics the model must generate; if two flows with different turbulent kinetic energies and integral length scales share the same Reynolds number, the conditioned samples cannot distinguish them.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model cuts inlet turbulence storage to 32 MB","AI generates turbulent inflow without bulky libraries","One diffusion model replaces entire turbulence database","Synthesize inlet turbulence on the fly with diffusion","32 MB model churns out realistic turbulence for CFD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1186,"prompt_tokens":723,"completion_tokens":463,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":467,"completion_tokens_details":{"reasoning_tokens":392}},"tokens_in":467,"tokens_out":463,"duration_ms":5744,"temperature":1.0,"reasoning_tokens":392,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T00:42:34.710940+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate two DHIT boxes with the same $Re_{L_{\\mathrm{int}}}$ but different pairs $(k, L_{\\mathrm{int}})$ — for example, double $k$ and halve $L_{\\mathrm{int}}$ so the product is unchanged — and condition the model at that Reynolds number. If the model produces the same energy spectrum and integral length scale for both, the conditioning is degenerate and the central premise fails.","supporting_citations":[],"review_version":1}