{"id":"4a8cae31-7822-4b48-8760-4947234bf7eb","arxiv_id":"2608.05208","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Strain rate outperforms vorticity and vortex deformation in forecasting extreme local uncertainty-production events at low Reynolds numbers, but all these fields fail near average strain and vorticity values.","lead":"This paper uses committor functions to estimate, from direct numerical simulations, the probability that a patch of turbulent flow will soon experience an extreme burst of uncertainty production. It finds that strain rate is the best single-flow predictor at low Reynolds numbers, and that regions where strain and vorticity are near their averages resist stable probabilistic forecasts.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Committor trained on box-averaged fields is applied to pointwise fields in Fig. 13, so the claimed unpredictable near-average region may be a scale-mismatch artifact.","rationale":"The reader's weakest assumption was the Markovian reduced dynamics and representativeness of analogue states; I agree that representativeness is central, but I identify a more specific and testable failure mode: the training and testing states are not drawn from the same phase-space distribution (box-averaged versus pointwise fields). This directly affects Figure 13 and the abstract's claim about an unpredictable near-average region, which is one of the paper's two headline results. The other headline result, the low-Reynolds-number Brier-score ranking of strain rate over vorticity, is based on scale-consistent box-averaged evaluation and is less threatened by this concern. The paper deserves credit for validating the committor against direct empirical probabilities in Figure 10 and for reporting the Reynolds-number dependence honestly; those elements are not in question. The proposed check is inexpensive: rerun the same pipeline at a single consistent scale, either pointwise or box-averaged, and compare the qualitative structure of Figure 13. Because the paper is already conditional and this concern is addressable without new physics, I keep the verdict at CONDITIONAL rather than moving to REJECT or ACCEPT.","tokens_in":37529,"tokens_out":7040,"duration_ms":77290,"concrete_test":"Retrain the analogue Markov chain using pointwise (single grid-point) predictor samples, preserving the same total number of states and spatial decorrelation requirements, and recompute the pointwise committor map of Figure 13. Alternatively, recompute Figure 13 using box-averaged testing boxes (the same scale as the training data and Brier-score evaluation). If the near-average 'unpredictable' region persists under scale-consistent training and testing, the claim is robust; if it disappears or changes qualitatively, the original figure is an artifact of the box-average-to-pointwise mismatch.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 3.2 defines analogue states as pointwise predictor vectors in R^D, but Section 4.3.1 states that the analogue Markov chain is constructed from box-averaged observed events, <Z_n>_D, over observation boxes of size 2L_D. The Brier-score evaluation in Section 4.3.4 consistently uses the same box-averaged testing boxes, so the predictor ranking (strain rate best at low Re) is internally consistent. However, Section 6.2 and Figure 13 compute the committor 'at every grid point of the DNS' using pointwise values of ||Omega||_F and ||S||_F, interpolated via the nearest-neighbor formula (3.12) over the box-averaged training states. Box-averaged fields have reduced variance, smoother time evolution, and different autocorrelation than pointwise fields; the transition statistics learned from one scale need not apply to the other. Pointwise states near the spatial mean will typically have few close analogues among box-averaged states, so the K-nearest-neighbor average in (3.12) may mix distant, unrepresentative analogues, producing the noisy, rapidly varying committor that is interpreted as 'probabilistic unpredictability' in the near-average region. Thus the abstract's second headline claim, that stable probabilistic forecasts are impossible where strain rate and vorticity are near their spatial averages, rests on a scale mismatch between training and testing data. If the committor were trained and evaluated at the same scale, the unpredictable region could disappear or move, directly weakening the central claim. This is a concrete, internal-consistency issue rather than a disagreement with turbulence consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript derives an evolution equation for the local uncertainty-production term P_Delta from the three-dimensional incompressible Navier-Stokes equations, identifies strain rate, vorticity, and vortex deformation as candidate local predictors, and estimates committor functions from DNS data using an analogue Markov chain. The predictor ranking is evaluated with Brier scores in four flow configurations (two forcing types, two Reynolds numbers each). The two headline claims are that at lower Reynolds numbers the strain rate is the best single-flow-field predictor of future extreme P_Delta events while vorticity has weak predictive power, and that near the spatial means of strain rate and vorticity the committor varies so rapidly that stable probabilistic forecasts are effectively impossible.","tokens_in":37856,"tokens_out":5117,"duration_ms":55557,"significance":"If the claims hold, the Brier-score ranking identifies a cheap single-flow precursor for localized uncertainty bursts, and the near-average unpredictability region would be an important practical caveat for reduced-order forecasting of turbulence. The paper has clear strengths: a complete derivation of the P_Delta evolution equation from Navier-Stokes, four systematically varied DNS configurations, multiple perturbation realizations for each case, a transparent analogue-Markov-chain methodology, and a direct comparison between committor estimates and empirical probabilities for single predictors in Figure 10. The central statistical framework is well suited to the question, and the paper is a plausible step toward probabilistic local predictability of turbulent uncertainty production.","major_comments":[{"comment":"The committor is trained and Brier-tested on box-averaged fields, as stated in Section 4.3.1 and used consistently in Section 4.3.4, but Figure 13 evaluates the committor at every grid point using pointwise values of the predictors and the nearest-neighbor interpolation formula (3.12). Box averaging reduces variance and changes the temporal autocorrelation of the fields, so pointwise states near the spatial mean may have few close analogues among the box-averaged training states, and the K-nearest-neighbor average in (3.12) may then mix distant, unrepresentative analogues. The rapid committor fluctuations interpreted as probabilistic unpredictability in the near-average region could therefore be a scale-mismatch artifact. Please recompute the Figure 13 field from box-averaged testing boxes, or retrain the committor on pointwise fields, before claiming that one-point kinematic fields cannot resolve predictability near the average.","section":"Section 6.2, Figure 13; Sections 4.3.1 and 4.3.4; Eq. (3.12)"},{"comment":"The analogue Markov chain assumes that the reduced predictor process is Markovian with time step Delta t, an assumption stated in Section 3.1 and used to build the chain in Sections 3.2 and 3.3. This assumption is not tested. If the true reduced dynamics has memory longer than Delta t, the committor computed from Eq. (3.8) is biased and the Brier-score ranking of predictors could change. The agreement with directly sampled empirical probabilities in Figure 10 validates the single-step conditional distribution but does not validate the multi-step Markov property on which the committor equation rests. Please add a quantitative memory check, for example comparing transition probabilities conditioned on one-step versus two-step histories, or reporting the autocorrelation decay of the predictor fields relative to Delta t.","section":"Section 3.1 and Section 3.3"},{"comment":"The analysis drops the third-order term -R_ij w_k d_k S_ij on the basis of |w| << |u|, and then focuses on the inertial contribution (i), dropping the pressure contribution (ii), the viscous contribution (iii), and the forcing contribution (iv) without a quantitative estimate of their magnitudes. Since the choice of strain rate, vorticity, and vortex deformation as predictors is motivated by the retained inertial terms, the theoretical basis for the predictor ranking would be substantially strengthened by a DNS-based estimate of each term in Eq. (2.4) within the exponential-growth regime. Such an estimate would show whether the neglected pressure and viscous contributions could change the relative importance of the candidate predictors.","section":"Section 2.1, Eq. (2.4)"},{"comment":"The headline ranking at lower Reynolds numbers -- strain rate best, vorticity worst -- rests on differences between Brier-score curves, but the text reports no statistical significance test for these differences. The standard-deviation bands shown in Figures 8 and 9 can be comparable to the separation between curves, particularly at intermediate values of alpha. Please provide a significance test, for example a paired bootstrap over testing boxes or over perturbation realizations, and state whether the ranking is statistically robust at each value of alpha and for each case.","section":"Section 6.1, Figures 8 and 9"}],"minor_comments":[{"comment":"In the sentence 'measure the precision of that estimation using the Brier score in ref 3.4', the cross-reference should read 'Section 3.4' rather than 'ref 3.4'.","section":"Section 3, introductory paragraph"},{"comment":"The typesetting of Eq. (3.14) contains a stray vertical bar and mismatched parentheses in the expression involving the square root; the formula should be displayed so that the decomposition into the irreducible variance term and the squared error term is unambiguous.","section":"Eq. (3.14)"},{"comment":"The word 'normalsed' appears in both captions and should be 'normalised'.","section":"Figure 11 and Figure 12 captions"},{"comment":"The word 'probablities' should be 'probabilities'.","section":"Section 6.2, final paragraph"},{"comment":"The word 'alignement' should be 'alignment'.","section":"Conclusion, last paragraph before Acknowledgments"},{"comment":"The caption contains a double comma after 'F2 512' in the listing of panels.","section":"Figure 7 caption"},{"comment":"The 'Key words' section is empty; either provide keywords or remove the heading.","section":"Abstract and front matter"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of JFM and the central statistical approach is appropriate. The main obstacle is the scale mismatch in Figure 13, which directly supports the abstract's second headline claim; this must be resolved. The Markovianity and neglected-term concerns also call for quantitative checks, but they are addressable within the manuscript's scope. No issues with novelty or citation practice arose in my review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this paper has one genuinely new analytical result (evolution equation for local uncertainty production, Eq. 2.4) and one genuinely new empirical application (committor functions, estimated with analogue Markov chains, for extreme uncertainty-production events that can occur anywhere in a turbulent flow). The Q-R plane analysis is standard, but the Brier-score ranking of predictors is new, and the authors deserve credit for reporting that vorticity increases the probability of extreme events even though the theory suggests it should decrease it. That kind of honesty is rare.\n\nThe empirical core—strain rate is the best single-flow predictor at low Reynolds number, vorticity is weak—is based on Brier scores computed on box-averaged testing boxes, so it is internally consistent. The training and testing are at the same scale. My main concern is with the second headline claim: the region near the spatial averages where \"even probabilities are effectively unpredictable.\" Figure 13 computes the committor at every grid point using pointwise values of ||S||_F and ||Omega||_F, but the analogue Markov chain was trained on box-averaged fields. Box averages have lower variance and smoother time evolution than pointwise fields. A pointwise state near the spatial mean will have few close analogues among box-averaged training states, so the K-nearest-neighbor average can mix distant, unrepresentative states and produce exactly the noisy, rapidly varying committor that is interpreted as probabilistic unpredictability. This is a scale mismatch, not a physical conclusion. The authors need to validate the pointwise interpolation or retrain the committor at the pointwise scale.\n\nTwo smaller issues. The derivation of Eq. (2.4) drops pressure and viscous terms after the initial inertial-term analysis, with qualitative justification but no quantitative check on the DNS data. A quick numerical estimate of the relative magnitudes would close that gap. And the Markovianity assumption for the coarse-grained dynamics is asserted but not tested; the committor could be biased if memory is longer than Delta t, and sensitivity to Delta t is not reported.\n\nWho is this for? Turbulence predictability researchers and anyone using committor methods in fluid dynamics. It deserves a serious referee. The scale-mismatch issue is concrete and fixable, and the analytical part is a real contribution. I would ask the authors to retrain at the pointwise scale and re-examine the Figure 13 claim before publication. My guess is the strain-rate ranking survives; the unpredictable-region claim may not.","headline":"New evolution equation and honest statistics, but the claim of an unpredictable near-average region likely rests on a scale mismatch between box-trained committors and pointwise inputs.","tokens_in":38375,"tokens_out":3536,"would_cite":true,"duration_ms":37995,"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":"In lower-Reynolds-number turbulence, strain rate is the best single-field predictor of extreme uncertainty production, while vorticity is weak and near-average states are unpredictable.","keywords":["turbulence predictability","uncertainty production","committor function","analogue Markov chain","strain rate","vorticity","Brier score","velocity gradient invariants"],"falsifier":"Run the same committor training on a fresh DNS realization at the same Reynolds number and compare out-of-sample Brier scores for $\\|S\\|_F$ and $\\|\\Omega\\|_F$: the central claim fails if vorticity matches or beats strain rate at low Reynolds number, or if adding a nonlocal predictor such as pressure or two-point velocity correlations removes the near-average unpredictability band.","tokens_in":37317,"feed_emoji":"🌪️","tokens_out":10015,"duration_ms":88238,"temperature":0.7,"pith_summary":"This paper asks which local flow quantities reveal, before it happens, that a small uncertainty in a turbulent flow will be violently amplified. It derives an evolution equation for the local uncertainty-production term from the Navier-Stokes equations, then estimates the probability of an extreme production event as a function of candidate predictor fields using a committor function built from an analogue Markov chain. In the lower-Reynolds-number simulations, the magnitude of the strain rate is the most informative single-flow predictor and vorticity is the weakest; at higher Reynolds numbers the ranking flattens. The paper also identifies a region near the spatial averages of strain rate and vorticity where even probabilistic forecasts are effectively impossible with these predictors. If this holds, local strain-rate measurements could serve as an early-warning diagnostic for where turbulence will decorrelate fastest, and forecasters should distrust predictions made near the average state.","feed_headline":"Strain rate beats vorticity as turbulence uncertainty predictor","feed_subtitle":"DNS-based committor shows local strain rate flags extreme uncertainty production, while near-average flow states stay unpredictable.","key_machinery":"The central object is the committor function $q(z_0) = P(T_B < T_A \\mid z(0)=z_0)$, the probability that a coarse-grained flow state reaches the extreme positive production set $B=\\{P_\\Delta>1\\}$ before the normal negative-production set $A=\\{P_\\Delta<0\\}$. The machinery that makes it computable is the analogue Markov chain: sampled box-averaged predictor states $\\{Z_n\\}$ define a discrete phase space, each state can jump with probability $1/K$ to the successor of one of its $K$ nearest neighbours, the sets $A$ and $B$ are made absorbing, and the committor is obtained from the eigenvectors of the resulting transition matrix by solving the linear system (3.9)-(3.11). The Brier score separates the intrinsic probabilistic floor from the error of the approximate committor, so comparing scores across predictor choices measures how much information each predictor retains about future extremes.","core_discovery":"This paper studies the local production of uncertainty in forced three-dimensional Navier-Stokes turbulence, defined as $P_\\Delta = -w_i S_{ij} w_j$, where $w$ is the difference between two nearly identical velocity fields and $S$ is the strain-rate tensor of the reference flow. From the Navier-Stokes equations it derives an evolution equation for $P_\\Delta$ (Eq. 2.4), identifying the inertial contributions of strain rate, vorticity, and vortex deformation. To forecast extreme positive $P_\\Delta$ events before they occur, the paper estimates the committor function --- the probability that a state reaches the extreme-production set $B$ before the normal set $A$ --- from DNS data using an analogue Markov chain, and scores the resulting probabilistic forecasts with the Brier score. It finds that in lower-Reynolds-number cases the strain-rate magnitude $\\|S\\|_F$ has the lowest Brier score among single-flow predictors, while vorticity $\\|\\Omega\\|_F$ has weak predictive power; at higher Reynolds numbers the three velocity-gradient predictors become nearly indistinguishable. It also finds a band of states near the spatial-average values of strain rate and vorticity where the committor varies rapidly and no stable probabilistic forecast can be made from these one-point fields.","pith_inferences":["Beyond the paper: if strain rate is the dominant precursor, velocity-gradient measurements in experimental or operational flows could be processed in real time to issue probabilistic warnings of uncertainty bursts, a step the paper does not take.","Beyond the paper: the near-average unpredictability band is probably a signature of missing state variables (pressure, alignment, two-point structure); a testable extension is to add the pressure Hessian or a second-point strain correlation as a predictor and see whether the band shrinks.","Beyond the paper: the same pipeline could rank precursors for other intermittent extremes in turbulence --- local dissipation, enstrophy bursts, or Reynolds-stress events --- and would reveal whether strain-rate dominance is specific to uncertainty production or generic.","Beyond the paper: the box-size dependence of the Brier ranking is untested; reducing $L_D$ toward the Taylor scale should sharpen or wash out the strain-rate advantage, and that would be a clean numerical experiment."],"forward_implications":["A single measured field --- the local strain-rate magnitude --- can flag where uncertainty will be produced, without needing to know the perturbed flow simultaneously.","Forecast systems for turbulence should either avoid, or supplement with extra information, regions where strain rate and vorticity sit near their spatial averages.","At higher Reynolds numbers, no one-point velocity-gradient scalar dominates; predictive skill depends on the alignment of uncertainty with strain and on fuller flow information.","The $Q$-$R$ topology maps extreme uncertainty production onto known strain self-amplification and vortex-stretching regions, giving testable spatial signatures of future decorrelation.","The analogue-Markov-committor method transfers to other localized extreme events anywhere in a turbulent flow, not only events at fixed positions or windows."],"supporting_citations":[{"why":"Derives the evolution equation for mean uncertainty energy, defines $P_\\Delta$, and supplies the DNS perturbation setup and self-similar PDF statistics that justify the sampling.","marker":"Ge et al. (2023)"},{"why":"Supplies the analogue Markov chain method for estimating committor functions from data and the Brier-score evaluation used throughout.","marker":"Lucente et al. (2022b)"},{"why":"Provides the spectral eigenvector procedure for solving the committor equation with absorbing sets.","marker":"Prinz et al. (2011)"},{"why":"Introduces the analogue method for forecasting in chaotic systems that the analogue Markov chain generalizes.","marker":"Lorenz (1969b,c)"},{"why":"Defines the Brier score used to measure forecast skill and to compare predictor choices.","marker":"Brier (1950)"},{"why":"Defines the eddy, convergence, and streaming regions from $Q$ that structure the flow-topology analysis.","marker":"Hunt et al. (1988)"},{"why":"Establishes the strain self-amplification tail $(27/4)R^2+Q^3=0$ used to locate extreme uncertainty production in the $Q$-$R$ plane.","marker":"Vieillefosse (1982, 1984)"},{"why":"Frames the committor as the solution of the backward Chapman-Kolmogorov equation, the starting point of the numerical scheme.","marker":"E & Vanden-Eijnden (2006)"}],"fun_headline_variants":["Strain rate beats vorticity for turbulence uncertainty at low Reynolds","Near-mean strain and vorticity states foil turbulence forecasts","Reynolds number changes which flow field predicts turbulence uncertainty","Turbulence uncertainty prediction fails near average strain, vorticity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole forecast construction assumes that the box-averaged, normalized predictor state evolves as a Markov chain with time step $\\Delta t$, so that the analogue neighbours and transition probabilities trained from DNS are a faithful representation of the flow's future; if the reduced dynamics has longer memory or the analogues do not represent the flow, the committor is biased and the predictor ranking could change.","fun_headline_variants_meta":{"raw":{"variants":["Strain rate beats vorticity for turbulence uncertainty at low Reynolds","Near-mean strain and vorticity states foil turbulence forecasts","Reynolds number changes which flow field predicts turbulence uncertainty","Turbulence uncertainty prediction fails near average strain, vorticity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000597,"raw_usage":{"total_tokens":2828,"prompt_tokens":1014,"completion_tokens":1814,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":1745}},"tokens_in":630,"tokens_out":1814,"duration_ms":14391,"temperature":1.0,"reasoning_tokens":1745,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T18:14:16.599494+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same committor training on a fresh DNS realization at the same Reynolds number and compare out-of-sample Brier scores for $\\|S\\|_F$ and $\\|\\Omega\\|_F$: the central claim fails if vorticity matches or beats strain rate at low Reynolds number, or if adding a nonlocal predictor such as pressure or two-point velocity correlations removes the near-average unpredictability band.","supporting_citations":[{"cited_title":", Held, M","cited_arxiv_id":null,"evidence_quote":"Provides the spectral eigenvector procedure for solving the committor equation with absorbing sets."},{"cited_title":"Monthly weather review 78 (1), 1--3","cited_arxiv_id":null,"evidence_quote":"Defines the Brier score used to measure forecast skill and to compare predictor choices."},{"cited_title":", Wray, A","cited_arxiv_id":null,"evidence_quote":"Defines the eddy, convergence, and streaming regions from $Q$ that structure the flow-topology analysis."},{"cited_title":"1982 Local interaction between vorticity and shear in a perfect incompressible fluid","cited_arxiv_id":null,"evidence_quote":"Establishes the strain self-amplification tail $(27/4)R^2+Q^3=0$ used to locate extreme uncertainty production in the $Q$-$R$ plane."}],"review_version":1}