{"id":"da13aa17-68b1-478f-b7c9-6e03e3b4db67","arxiv_id":"2505.02580","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A PINN with a prescribed mean temperature profile reconstructs temperature and pressure fields from Lagrangian velocities with about 90% correlation in hard-turbulence Rayleigh-Bénard convection, using both DNS and experimental PTV data.","lead":"A physics-informed neural network trained on Lagrangian particle velocities reconstructs temperature and pressure fields in turbulent Rayleigh-Bénard convection at Rayleigh number 10^9. The method is validated against direct numerical simulation and applied to experimental particle tracking velocimetry, offering a path to full velocity, temperature, and pressure measurements in convective flows without direct thermometry.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The prescribed mean-temperature profile in Eqs. (6)-(7) is internally inconsistent as printed, so the central calibration a=2Nu is not supported by the written derivation.","rationale":"The reader correctly identified the prescribed mean temperature profile as the weakest assumption, focusing on the transfer from DNS to the experimental cell with sidewall heat loss. My stress-test found an even more direct problem: the profile as written in Eq. (7) is internally inconsistent with Eq. (8) and with the temperature boundary conditions. For a=126, the printed formula gives Tmean≈+0.5 at both plates and in the bulk, and the wall derivative is essentially zero, so Eq. (8) cannot be derived. This affects the central DNS validation, not just the experimental extrapolation. I do not accuse the authors of dishonesty; this may be a typographical error in the manuscript, and the code is available for verification. But as written, the central derivation is not reproducible. Since the paper's core result may still be correct with the intended formula, a conditional verdict remains appropriate, now with an explicit mandatory check of Eq. (7)-(8) and the actual code. I therefore leave the reader's verdict unchanged while strengthening the condition.","tokens_in":13874,"tokens_out":17412,"duration_ms":199288,"concrete_test":"Evaluate Eq. (7) at Z=0, Z=0.5, and Z=1 for a=126, and compute dTmean/dZ at Z=0; verify that Tmean(1) = -0.5 and that -dTmean/dZ at the wall equals Nu = 63 via Eq. (8). Also check whether a=1 gives a linear profile as claimed. If any of these checks fail, inspect the GitHub implementation to identify the actual formula used; if the implementation matches Eq. (7), the central calibration and the resulting temperature/pressure reconstructions are unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The method's core mechanism is the empirical mean temperature profile Tmean(Z,a) in Eq. (6)-(7), which fixes the vertical mean temperature and the temperature boundary conditions, and is calibrated by setting a=2Nu via Eq. (8). As printed, Eq. (7) does not satisfy the stated boundary conditions: for a=126, e^{-0.5a} is essentially zero, so both branches reduce to Tmean(Z)≈+0.5 for all Z in [0,1]. The top plate therefore has T=+0.5 instead of the required -0.5, and the well-mixed bulk is not near zero. Moreover, differentiating the lower branch at Z=0 gives dTmean/dZ = 2a e^{-0.5a}, which for a=126 is essentially zero, so the central relation Nu = a/(2 - 2e^{-0.5a}) ≈ a/2 stated in Eq. (8) does not follow from Eq. (7). The text's claim that a=1 yields a linear temperature profile is likewise inconsistent with the printed expression. This is a load-bearing problem: the DNS validation in Sec. 4 and the experimental transfer in Sec. 5 both rely on this profile, and the reported 90% correlation cannot be reproduced or trusted from the equations as written. If the GitHub implementation uses a different formula than Eq. (7), the paper must state that formula; if it uses the printed formula, the calibration is invalid. This concern is more fundamental than the reader's transferability concern, because it affects the derivation itself.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a physics-informed neural network (PINN) method to reconstruct temperature and pressure fields from Lagrangian velocity data in turbulent Rayleigh-Bénard convection at Ra=10^9 and Pr=6.9. The network, a multilayer perceptron with sine activation, is trained on DNS-generated particle tracks (150,000 particles) and on experimental PTV tracks (about 50,000 active particles per time step). A key ingredient is an empirical mean temperature profile Tmean(Z,a) with parameter a set to 2Nu, which imposes the wall boundary conditions and reduces the difficulty of the near-wall gradients. In the DNS validation, the reconstructed temperature and pressure reach Pearson correlations of about 90% and mean absolute errors around 10^-2 and 10^-3, respectively; the reconstructed Nusselt number is 61.8 compared with 63.4 in the DNS. For the experiment, the same a=126 is used, yielding a Nusselt number of 34.6, later corrected to 55.3 using a velocity-scale factor. The paper claims the method can provide temperature and pressure fields in the hard-turbulence regime from PTV data alone, and it makes the PINN code available open-source. I also checked the internal consistency of Eq. (7): the printed mean-temperature profile does satisfy Tmean(0)=+0.5, Tmean(1)=-0.5, and yields Eq. (8) by differentiating at Z=0, so the stress-test concern about Eq. (7) does not land.","tokens_in":14206,"tokens_out":9572,"duration_ms":105857,"significance":"If the result holds, the method would be a practical step toward measuring all three key fields (velocity, temperature, pressure) in turbulent thermal convection using only particle-tracked velocity data. The DNS validation is quantitative, with high correlation on temperature and pressure, and the open-source code is a strength for reproducibility. The use of a prescribed mean profile is a reasonable way to overcome the hard-turbulence difficulty, and the experimental demonstration is a useful proof of concept, though it lacks ground-truth temperature or pressure data. The main value of the paper lies in showing that the approach works in a regime (Ra=10^9) that previously had only been demonstrated at lower Rayleigh numbers.","major_comments":[{"comment":"The PCC and MAE values are computed on the last 10 time steps of the DNS dataset, which are part of the 125 time steps used for training the PINN through the data-loss term. Because the manuscript does not describe a held-out validation partition, the 90% correlation does not establish that the PINN can reconstruct temperature and pressure from velocity data outside the training interval. Please either train on the first 115 time steps and evaluate on the final 10, or explicitly state that the reported correlation is a training-fidelity metric rather than a generalization metric.","section":"Section 4, evaluation metrics (Fig. 8)"},{"comment":"The mean temperature profile parameter a is set to 2Nu with Nu=63.4 taken from the DNS, so the reported Nu_PINN=61.8 is largely constrained by the prescribed mean profile and is not an independent physical validation. The manuscript should present this as a consistency check. Additionally, reporting the correlation and MAE for the temperature fluctuation field T' alone, which is the part actually learned by the network, would strengthen the validation and clarify how much of the 90% correlation is attributable to the prescribed mean profile.","section":"Section 4, Nusselt-number validation"},{"comment":"In the experimental case there is no measured temperature or pressure ground truth, and the mean temperature profile is prescribed with a=126 based on the DNS Nusselt number. The reconstructed Nu_PINN=34.6 deviates strongly from the assumed value of 63, and the correction factor of 1.6 is derived solely from the ratio of maximum vertical velocities, not from an independent heat-flux measurement. The manuscript should either obtain or cite an experimental Nusselt number for the cell, or explicitly label the experimental temperature and pressure fields as model-dependent reconstructions whose accuracy is not yet validated.","section":"Section 5, experimental reconstruction"}],"minor_comments":[{"comment":"The text states the maximum observed velocity is about ±10 m/s, which is inconsistent with the later statement that the maximum normalized vertical velocity is about 0.18 in the experiment (with a reference velocity of about 50 mm/s). The units appear to be a typo and should be corrected (e.g., ±0.10 m/s or ±10 cm/s).","section":"Section 2.2, paragraph on track visualization"},{"comment":"The statement that a=1 results in a linear temperature profile is only approximately true; the profile is exponential, and for a=1 the deviations from a straight line connecting the boundary values are small but visible. Consider rephrasing to 'approximately linear' or 'close to linear'.","section":"Section 3.1, discussion of Eq. (7)"},{"comment":"The factor (Z^2 - Z) multiplying the temperature fluctuation is introduced without explanation. It would be helpful to state explicitly that this factor forces the fluctuation to vanish at the top and bottom plates, so that the boundary conditions are exactly satisfied by Tmean alone.","section":"Section 3.1, Eq. (6)"},{"comment":"The caption says the profile is 'predicted in the DNS', but the figure shows the empirical function Tmean(Z,a) for increasing a. Clarify that the DNS data are used to motivate the value a=126, while the curves are from Eq. (7).","section":"Figure 7 caption"},{"comment":"The reference to Barta et al. (2023) is marked 'under review'; if a published version is available, it should be updated.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The central derivation via Eq. (7) is sound; the stress-test concern about internal inconsistency appears to be based on a misreading of the exponential terms. The main weaknesses are methodological: the DNS validation is performed on training data, the Nusselt-number agreement is partly circular, and the experimental transfer is unvalidated. These can be addressed with a re-analysis or a clearer framing of the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know two things about this paper before you read it. First, it is a credible extension of the authors' earlier PINN work: they now tackle Ra=10^9 hard-turbulence RBC and apply the method to real PTV data, with a DNS validation that mostly holds up. Second, the main weakness is not the math of the mean-temperature profile — a stress-test worry about Eq. (7) does not survive close reading — but a circularity in how the profile parameter is set.\n\nThe new results are real: 150,000 DNS Lagrangian tracers and 50,000 active tracks per step in a water-filled cell; the PINN with a prescribed mean profile plus learned fluctuations reconstructs T and p with about 0.9 Pearson correlation and sensible MAEs. The mean-profile trick (Eq. 6-8) is a reasonable way to handle the steep wall-to-bulk temperature drop, and the implementation is self-consistent. The code is promised as open source and compatible with proPTV. The authors are also candid about boundary-layer failures and the sidewall heat loss in the experiment.\n\nThe soft spots are proportionate. The Nusselt-number check in the DNS (Nu_PINN=61.8 vs 63.4) is not an independent confirmation, because a is set to 2Nu from the very same DNS. The high temperature correlation is likewise aided by prescribing the mean profile; the genuinely learned part is the fluctuation field. That still counts, but it lowers the significance of the 90% headline. The experimental section has no ground truth for T or p, so \"realistic\" is an inference from plausibility. The reported Nu=34.6 and the ad hoc 1.6 correction to get 55.3 are not convincing — the heat-loss explanation is plausible but unmeasured. No error bars or uncertainty estimates appear anywhere.\n\nThe stress-test note about Eq. (7) being inconsistent is, as far as I can tell from the printed equations, wrong. The profile satisfies T(0)=0.5, T(1)=-0.5, is continuous at Z=0.5, and differentiating the lower branch at Z=0 gives Nu=-dT/dZ=a/(2-2e^{-a/2})≈a/2, which is exactly Eq. (8). The only sloppy sentence is the claim that a=1 gives a linear profile; that is an approximation, not an exact statement, and it has no bearing on the results.\n\nWho is this for? Anyone working on PINN-based reconstruction in convection, and experimentalists who want an estimate of T and p from PTV without adding thermometry. It deserves a serious referee. I would send it to review with the expectation of heavy revision: ask for a sensitivity study in a, a baseline comparison (e.g., a=1 or a fitted value), and a clear statement that the experimental part is a feasibility demonstration, not a validated measurement.","headline":"A credible extension of PINN-based temperature/pressure reconstruction to hard-turbulence RBC, but the DNS-calibrated mean profile makes the headline validation less independent than it looks.","tokens_in":14767,"tokens_out":6692,"would_cite":true,"duration_ms":67114,"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 physics-informed neural network reconstructs temperature and pressure from Lagrangian velocity data in turbulent Rayleigh-Bénard convection, matching DNS ground truth at 90% correlation.","keywords":["physics-informed neural networks","Rayleigh-Bénard convection","particle tracking velocimetry","temperature reconstruction","pressure reconstruction","Lagrangian velocity data","data assimilation","turbulent convection"],"falsifier":"Place a vertical rake of fine thermistors in the experimental cell, time-average their readings over many large-scale-circulation periods, and compare the resulting mean temperature profile with $T_{\\mathrm{mean}}(Z, a=126)$. If the measured profile deviates beyond boundary-layer uncertainty, the reconstructed temperature field and derived heat-transport number are set by the assumed prior rather than by the data.","tokens_in":13622,"feed_emoji":"🌡️","tokens_out":9330,"duration_ms":98771,"temperature":0.7,"pith_summary":"This paper tries to establish that a single physics-informed neural network, fed only Lagrangian velocity measurements from tracer particles, can reconstruct the temperature and pressure fields of turbulent Rayleigh-Bénard convection at $Ra = 10^9$ and $Pr \\approx 7$. If true, it would give experimentalists a way to obtain all three main flow variables from particle tracking alone, without separate thermometry or pressure probes. The claim is demonstrated two ways: against DNS ground truth, where reconstructed fields reach about 90% correlation, and on experimental PTV data, where the network produces physically realistic plumes, large-scale circulation, and pressure extrema while also smoothing noisy velocity components.","feed_headline":"PINN turns particle tracks into temperature and pressure fields","feed_subtitle":"In hard-turbulence Rayleigh-Bénard convection, reconstructed fields match DNS ground truth at 90 percent correlation.","key_machinery":"The central object is a modified PINN output: the MLP emits a temperature fluctuation $T'$ and pressure $p$, and the true temperature is assembled as $T = T_{\\mathrm{mean}}(Z,a) + (Z^2 - Z)T'$, where $T_{\\mathrm{mean}}(Z,a)$ is an empirical two-sided exponential vertical profile. This form enforces the top and bottom temperature boundary conditions exactly and converts the hard part of the problem, steep wall gradients and a well-mixed bulk, into a fluctuation-learning task. The parameter $a$ is fixed by the Nusselt number through $Nu = a/(2 - 2e^{-0.5a}) \\approx a/2$, set to $a = 126$ from the DNS value $Nu \\approx 63$. The rest of the machinery is a standard equation-constrained MLP training loop: Adam optimization of a loss combining velocity data, momentum and energy and divergence residuals, pressure centering, and velocity boundary conditions, with sine activation functions that help the network represent many wavenumbers.","core_discovery":"On the paper's own terms, the discovery is that hard-turbulence Rayleigh-Bénard convection ($Ra = 10^9$, $Pr \\approx 7$) is no longer out of reach for PINN-based assimilation: once the network is tasked with learning only temperature fluctuations around an empirical mean profile instead of the full temperature field, a single multilayer perceptron with sine activations can infer both temperature and pressure from Lagrangian velocity data. In the DNS validation this yields Pearson correlations around 0.90 for temperature and pressure at particle positions, with mean absolute errors near $10^{-2}$ for temperature and $10^{-3}$ for pressure. On experimental PTV data with the same nominal parameters, the same recipe produces thermal plumes, a large-scale circulation, and pressure extrema at the expected locations, and it visibly smooths measurement noise in the camera-direction velocity component.","pith_inferences":["Editorial inference: because $a$ is fixed from DNS rather than learned, the method is not fully data-driven; treating $a$ as a trainable parameter or inferring it from sparse temperature measurements would make the pipeline self-calibrating in setups with unknown heat loss.","Editorial inference: the 90% correlation is scored on particle positions, where boundary layers are absent; the paper's own error maps show the largest deviations near the heating and cooling plates, so an Eulerian grid-based validation including boundary layers would likely report a lower correlation.","Editorial inference: the same mean-profile trick should transfer to other wall-bounded convective flows whenever a reliable analytic mean-temperature form exists; flows without such a form would need a different boundary-condition prior.","Editorial inference: combining Lagrangian velocity with sparse point thermometry, such as a few thermistors or thermochromic particles, would provide a direct experimental check and could extend the method's accuracy beyond the reported bulk-focused metrics."],"forward_implications":["With only tracer-particle velocities spanning about 2.5 free-fall times, the network reproduces temperature and pressure structures at roughly 90% correlation with DNS ground truth at particle positions.","Experimental PTV data can be upgraded from velocity-only to velocity-temperature-pressure fields, revealing thermal plumes, the large-scale circulation, and pressure minima in corner and central circulations.","The equation-constrained velocity output acts as a denoiser, particularly along the camera axis where PTV errors are largest.","The reconstructed flow can be used to estimate the Nusselt number, yielding 61.8 for DNS (close to 63.4) and a heat-loss-reduced value for the experiment that rises to about 55 after the authors' 1.6 velocity-scale correction.","Since the network is released as open-source code and plugs into the tracking workflow used for the experiment, the pipeline can be applied to other convection experiments without new hardware."],"supporting_citations":[{"why":"Supplies the base MLP-PINN architecture and loss weighting that this paper extends to simultaneous temperature and pressure reconstruction.","marker":"Mommert et al., 2024"},{"why":"Establishes that roughly two free-fall times of Lagrangian data suffice and supplies the evaluation-metric choices used here.","marker":"Volk et al., 2025"},{"why":"Introduced the mean-temperature-profile prior that equations (6)-(7) generalize to the exponential form.","marker":"Toscano et al., 2024"},{"why":"Describes the cubic water-filled Rayleigh-Bénard cell and the PTV measurement procedure whose dataset is used in section 5.","marker":"Barta and Wagner, 2025"},{"why":"Provides the open-source particle-tracking software that delivers about 50,000 active tracks per time step in the experiment.","marker":"Barta et al., 2023"},{"why":"Supplies the a priori Nusselt estimate used for grid resolution and for setting the profile parameter $a = 2Nu$.","marker":"Grossmann and Lohse, 2000"},{"why":"Defines the PINN residual-based approach that this work applies to the Boussinesq equations.","marker":"Raissi et al., 2019"},{"why":"Defines the hard-turbulence regime ($Ra > 4 \\times 10^7$) that this paper targets.","marker":"Castaing et al., 1989"},{"why":"Provides the Batchelor-length resolution criterion used to size the DNS grid.","marker":"Scheel et al., 2013"}],"fun_headline_variants":["PINN from tracks: temperature and pressure in turbulent convection","PINN reconstructs thermal fields from particle tracks alone","90% correlation: PINN maps tracks to convection fields","Hard turbulence: PINN recovers T and P from tracks","PINN + particle tracks yields full thermal fields"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reconstruction leans on an assumed mean temperature profile calibrated to the simulation's heat-transport level remaining valid in the experimental cell, whose glass sidewalls lose about 20 percent of the heat; the much lower heat-transport estimate the network produces for the experiment (34.6 versus the assumed 63) suggests the premise is already strained.","fun_headline_variants_meta":{"raw":{"variants":["PINN from tracks: temperature and pressure in turbulent convection","PINN reconstructs thermal fields from particle tracks alone","90% correlation: PINN maps tracks to convection fields","Hard turbulence: PINN recovers T and P from tracks","PINN + particle tracks yields full thermal fields"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000808,"raw_usage":{"total_tokens":3569,"prompt_tokens":991,"completion_tokens":2578,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":2499}},"tokens_in":607,"tokens_out":2578,"duration_ms":21692,"temperature":1.0,"reasoning_tokens":2499,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:47:50.209301+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Place a vertical rake of fine thermistors in the experimental cell, time-average their readings over many large-scale-circulation periods, and compare the resulting mean temperature profile with $T_{\\mathrm{mean}}(Z, a=126)$. If the measured profile deviates beyond boundary-layer uncertainty, the reconstructed temperature field and derived heat-transport number are set by the assumed prior rather than by the data.","supporting_citations":[{"cited_title":"A PINN Methodology for Temperature Field Reconstruction in the PIV Measurement Plane: Case of Rayleigh-B\\'enard Convection","cited_arxiv_id":"2503.23801","evidence_quote":"Establishes that roughly two free-fall times of Lagrangian data suffice and supplies the evaluation-metric choices used here."},{"cited_title":"proptv: A probability-based particle tracking velocimetry framework","cited_arxiv_id":null,"evidence_quote":"Provides the open-source particle-tracking software that delivers about 50,000 active tracks per time step in the experiment."}],"review_version":1}