{"id":"e456dd06-52a0-4d22-b583-8071c33c99d5","arxiv_id":"2412.11266","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Bayesian inversion of mean flow MRI data jointly reconstructs a turbulent jet's mean velocity and infers the parameters of an algebraic eddy-viscosity model, demonstrated on an FDA nozzle phantom at Reynolds number 6500.","lead":"Flow MRI measurements of a turbulent jet in a nozzle are combined with physics-based RANS equations to jointly reconstruct the mean velocity field and learn the parameters of an effective viscosity model. The approach is demonstrated on a medical-device-like FDA nozzle, reducing the data-model mismatch, but it is not yet validated against independent measurements.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No independent validation supports the 'without overfitting' claim: the reported agreement is in-sample, the inferred inlet profile plus six-parameter mixing length can absorb model error, and posterior uncertainties are wide.","rationale":"The reader's verdict is CONDITIONAL, and this stress test finds the same core weakness: the Boussinesq algebraic closure is asserted rather than validated. The strongest claim—'successfully reconstructs the mean flow field and learns the most likely turbulence model parameters without overfitting'—is not supported by any independent check. My concern is slightly broader than the reader's weakest_assumption, because even if the algebraic closure happened to be adequate, the paper provides no cross-validation or held-out data to rule out overfitting through the inferred inlet boundary condition and the six-parameter mixing length. The reported discrepancies are training errors. The large posterior uncertainties reinforce this: parameters such as beta and xc are so weakly constrained that calling them 'learned' is generous. That said, this is a pilot proof-of-concept paper, and the Bayesian framework is standard; the issue is validation design, not internal inconsistency. The proposed hold-out test directly targets the weakest link and would either support or falsify the 'without overfitting' claim. I therefore do not change the reader's verdict: CONDITIONAL is appropriate, and the condition should be independent validation of the reconstruction and learned parameters.","tokens_in":7261,"tokens_out":3300,"duration_ms":34287,"concrete_test":"Mask a random 20% of the data voxels (or all voxels downstream of the jet breakdown region) and re-run the optimization in Sec. 4. Compute Eq. (4.1) on the held-out voxels. If held-out E/sigma exceeds the in-sample value by more than about 50%, the fit is overfitting and the learned parameters are artifacts; if it matches, the 'without overfitting' claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that the algorithm reconstructs the mean flow and learns the most likely turbulence parameters 'without overfitting'—depends entirely on in-sample agreement. In Sec. 4, the discrepancy (4.1) is evaluated on the same flow-MRI data assimilated during optimization; Figures 6(e,f) and the MAP discrepancies (1.30, 1.31, 2.08)/sigma therefore measure training error, not generalization. The model has substantial compensatory freedom: the inlet profile gi is inferred over the whole inlet, and the compound mixing-length parameters (beta, xc, c, ds0 in Table 3) are free, so the forward map Z in (2.1) can fit the data even if the Boussinesq closure (2.5)-(2.7) is wrong. The wide posterior uncertainties (beta = 19.2 +/- 12.6, xc = 3.25 +/- 2.87 cm) show the data barely constrain the turbulence parameters, so 'most likely parameters' is a statement about a Laplace approximation, not about the physics. The residual uz discrepancy of 2.08/sigma at the MAP additionally indicates structured model error or mis-specified sigma = 5 cm/s, violating the white-noise likelihood (2.2). Thus the advertised proof of concept is not established by the evidence presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript solves a Bayesian inverse RANS problem: it combines a finite-element RANS solver with an algebraic compound mixing-length eddy-viscosity closure (Eqs. 2.5–2.8) and uses adjoint-based gradient optimization plus Laplace's approximation (Eqs. 2.2–2.4) to infer the inlet boundary condition and six turbulence-model parameters from 3D flow-MRI data of a confined turbulent jet in the FDA nozzle at Re 6500. The authors report MAP estimates and uncertainties (Table 3), data-model discrepancies (Eq. 4.1, Figs. 5–6), and conclude that the algorithm reconstructs the mean flow and learns the most likely turbulence parameters 'without overfitting' (abstract; §4.1; §5).","tokens_in":7586,"tokens_out":5921,"duration_ms":53717,"significance":"If the central claims are fully supported, the paper would demonstrate a useful proof of concept: jointly reconstructing turbulent mean fields and calibrating a RANS closure from clinical-grade flow-MRI data, with quantified uncertainty, in a regime where laminar-assimilation tools fail. The experimental dataset (multi-VENC 4D-flow MRI, phantom construction, pressure ports) is a valuable contribution, and the algorithmic core—adjoint-accelerated Laplace-approximated Bayesian inference—is sound and clearly described. However, the validation presented is entirely in-sample and the parameter uncertainties are large; as it stands the paper establishes an interesting numerical pipeline but not the advertised 'without overfitting' generalization claim.","major_comments":[{"comment":"The claim of 'without overfitting' in the abstract and of a 'sufficiently descriptive' RANS model in §4.1 rests entirely on the data-model discrepancy (4.1), which is evaluated on the same flow-MRI data used to optimize gi and p. The MAP values (1.30, 1.31, 2.08)/σ are training residuals, not generalization errors. The model has ample capacity to fit those data even when the closure is wrong: gi is inferred over the full inlet and p contains six parameters (µℓ, α, β, xc, c, ds0). An independent check is required: hold-out voxels or velocity components, a synthetic-data twin, comparison with the measured pressure drop mentioned in §3.1, or another observable. The authors' own statement in §4.1 that the learned algebraic model 'is not expected to extrapolate well' further undercuts the blanket wording; the paper should be reframed as an in-sample assimilation proof of concept unless such a check is added.","section":"Abstract; §4.1, Eq. (4.1)"},{"comment":"At the MAP point the axial data-model discrepancy is 2.08σ (Sec. 4) with σ = 5 cm/s. Given the large number of voxels (40×45×169), a 2σ mean discrepancy is incompatible with the white-noise assumption ε ∼ N(0, σ²I) in Eq. (2.2), unless the residual is dominated by a small region. The paper neither plots the residual structure nor reports a spatial correlation test. This matters because the posterior covariance (2.4) is computed with that likelihood; if σ is misspecified the reported parameter uncertainties are unreliable. Please add residual diagnostics and either recalibrate σ or model correlated noise.","section":"§4, Eqs. (2.2), (4.1)"},{"comment":"The MAP uncertainties in Table 3 are extremely wide: β = 19.2 ± 12.6 and xc = 3.25 ± 2.87 cm, i.e., relative uncertainties of 66% and 88%. The claim that the algorithm 'learns the most likely turbulence model parameters' therefore overstates the information content. The posterior covariance in Eq. (2.4) is also a Laplace approximation around a possibly non-identified optimum; the paper should report the parameter correlation matrix or profile likelihoods, and dampen the conclusion to 'weakly constrained parameters with wide posterior.'","section":"§2, Table 3"}],"minor_comments":[{"comment":"The list of discrepancies contains typos: E(¯uy) appears twice and E(¯u◦y) uses inconsistent bar/circle notation; please list E(ux), E(uy), E(uz) with consistent notation.","section":"Eq. (4.1) and surrounding text"},{"comment":"The paper calls the model 'five-parameter' in several places, but p = (µℓ, α, β, xc, c, ds0) contains six parameters plus the inferred inlet profile gi; please correct the count.","section":"§2.1, Table 3"},{"comment":"The table header labels two columns as 'c [cm]'; one should be 'xc [cm]' and the other 'c [cm]'.","section":"Table 3"},{"comment":"The Gaussian noise assumption should be justified for the magnitude-weighted multi-VENC averaging described in §3.2, since MRI phase noise is commonly modeled as Rician or Rayleigh; either cite supporting reasoning or note that the Gaussian likelihood is an approximation.","section":"§3.2, Eq. (2.2)"},{"comment":"The discrepancy panels share the velocity colormap [0,100 cm/s], which compresses the residual structure; a symmetric or diverging colormap would better reveal where the model error actually remains.","section":"Figure 6(e,f)"}],"recommendation":"major_revision","confidential_remarks":"The paper is plausible as a workshop-proceedings contribution, but for a regular journal the central 'without overfitting' and 'sufficiently descriptive' claims need an independent validation step. The authors appear aware of the limitation in §4.1's extrapolation remark, so a revision adding a held-out-data or pressure-drop check would be a reasonable path. I would also ask for a data/code availability statement and a concrete check of the Laplace approximation's validity."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a legitimate proof of concept for inferring RANS model parameters and reconstructing a mean velocity field from flow MRI data in a turbulent confined jet. The novelty is the extension of the authors' earlier laminar and non-Newtonian Bayesian inversion work to a turbulent RANS setting with an algebraic eddy viscosity, and the FDA nozzle flow MRI dataset is a nice new test case. The experiment is carefully described, and the optimization convergence is shown. That part holds up.\n\nThe soft spot is the 'without overfitting' claim. The agreement shown in Figure 6(f) is in-sample: the discrepancy is measured on the very data used for optimization. The model has substantial compensatory freedom—the inlet profile gi is inferred as a full function, and the compound mixing length adds six parameters. With posterior uncertainties like beta = 19.2 ± 12.6 and xc = 3.25 ± 2.87 cm, the data barely constrain the turbulence parameters. The residual axial discrepancy of 2.08σ also suggests structured model error or misspecified noise, not the white Gaussian noise assumed in the likelihood. So the most likely parameters are a statement about a weakly identified Laplace mode, not about the physics. The stress-test note is on target.\n\nThat said, the authors are honest about the limitations: they explicitly say algebraic models are not expected to extrapolate well and flag the k-ε model as a next step. For a pilot study that is a reasonable posture. The central concept—that the Bayesian machinery can assimilate flow MRI data into a RANS problem and produce a plausible mean field—is still demonstrated, even if the specific parameter values are not reliable.\n\nThis paper is for people working on data assimilation in fluids, medical imaging, or turbulence model calibration. It doesn't validate the turbulence model, but it shows the inference pipeline working on a real turbulent flow MRI dataset. With a serious referee, the authors should be pushed to either add an independent validation (withheld data, a high-fidelity reference, or a predictive check) or substantially soften the overfitting claim. I'd send it to peer review rather than desk reject it; it's a useful data point, not a settled result.","headline":"A legitimate proof of concept for turbulent Bayesian RANS inversion from flow MRI, but the 'without overfitting' claim is in-sample only and should be softened.","tokens_in":8097,"tokens_out":2476,"would_cite":true,"duration_ms":23537,"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":"Bayesian inversion of the RANS equations reconstructs a turbulent jet's mean velocity field and learns its eddy-viscosity parameters directly from flow MRI data.","keywords":["Bayesian inversion","Reynolds-averaged Navier-Stokes equations","flow MRI","eddy viscosity model","turbulence model calibration","confined turbulent jet","FDA nozzle","Laplace approximation"],"falsifier":"Measure the full turbulent stress tensor in the same FDA nozzle at Reynolds number 6500 with particle image velocimetry, and compare it with the stress predicted by the inferred eddy-viscosity parameters; if the two disagree beyond measurement error, the inferred parameters are artifacts of the closure rather than true properties of the turbulence.","tokens_in":7105,"feed_emoji":"🌊","tokens_out":13154,"duration_ms":102872,"temperature":0.7,"pith_summary":"This paper aims to establish a proof of concept: a turbulent mean flow can be reconstructed and its turbulence model calibrated in one Bayesian pass, using only the coarse, noisy 3D velocity data that flow MRI provides. The authors solve an inverse Reynolds-averaged Navier–Stokes (RANS) problem in which the unknowns are the inlet boundary condition and the coefficients of an algebraic eddy-viscosity model; adjoint-based optimization under a Laplace approximation (a Gaussian fit to the posterior) finds the most probable parameters and their uncertainties. They test the method on a confined turbulent jet at Reynolds number 6500 in an idealized medical-device (FDA) nozzle, assimilating 4D flow MRI velocity data with 1 mm isotropic voxels. The reconstructed mean velocity field fits the measured data in both the jet-breakdown and pipe-flow regions, and the learned mixing-length parameters carry quantified error bars, which the authors read as evidence that the model is learning rather than overfitting. If this holds, it becomes possible to build RANS models whose turbulence closure parameters are informed directly by imaging data, without resolving turbulent eddies.","feed_headline":"Bayesian inversion learns turbulence model straight from flow MRI data","feed_subtitle":"The algorithm reconstructs the jet's mean velocity and learns eddy-viscosity parameters with quantified uncertainty.","key_machinery":"The load-bearing machinery is the Bayesian inverse RANS problem with adjoint-accelerated Laplace approximation. The forward map $Q$ sends unknown parameters $x = (g_i, p)$ — the inlet Dirichlet velocity and the turbulence-model parameters — to a RANS solution $u$; the projection $S$ maps $u$ into the data space, and the mismatch with the flow MRI data $u^\\star$ is measured under the Gaussian noise covariance $C_{u^\\star}$. The objective is the negative log-posterior of Eqs. (2.2)–(2.3), minimized with adjoint gradients, and the posterior is Laplace-approximated around the MAP point to give the covariance of Eq. (2.4). Inside $Q$, turbulence is closed with the Boussinesq ansatz $\\mu_t = \\ell_c^2 \\dot\\gamma$, where $\\dot\\gamma$ is the shear-rate magnitude and the compound mixing length is $\\ell_c = \\alpha H_\\eta(d_s - d_{s0}) + \\beta(1-H_\\eta)d_w$, with $H_\\eta$ a smooth activation that selects the jet-breakdown or pipe-flow regime. The parameters $p = (\\mu_\\ell, \\alpha, \\beta, x_c, c, d_{s0})$ are exactly what the inversion learns.","core_discovery":"The central claim is that a compact algebraic eddy-viscosity closure, built on a compound mixing length that switches between a streamwise-distance law in the jet-breakdown region and a wall-distance law in the pipe-flow region, is descriptive enough for this confined jet that its parameters can be learned from mean-velocity data alone. Solving the Bayesian inverse RANS problem yields a maximum-a-posteriori parameter set whose modeled velocity field matches the flow MRI data in both regions, with posterior covariances that place error bars on every inferred parameter. Because the data are noisy and the priors are broad, the authors interpret the match as successful learning of turbulence-model parameters rather than overfitting of a flexible model. The paper further claims the method is general: any differentiable turbulence model, algebraic or multi-equation, can replace the closure used here, and the same machinery extends to unsteady turbulent flows.","pith_inferences":["A sharp test the authors did not run: repeat the inference at a second flow rate in the same phantom and check whether the learned mixing-length parameters transfer; the paper itself warns that algebraic models are not expected to extrapolate well, so non-transfer would show the parameters are fitted constants rather than physical closure coefficients.","The local turbulent kinetic energy available from MR signal decay is an independent constraint the authors chose not to use; assimilating it would reduce the degeneracy between mixing-length amplitude and shear-rate distribution that a velocity-only fit may leave.","Because the posterior is only Laplace-approximated, a few MCMC samples around the MAP point would test whether the reported Gaussian error bars are reliable; if the posterior is skewed, the uncertainties on $\\beta$ and $x_c$ would need revision.","The phantom's two pressure ports are an untapped validation channel: the inferred RANS solution predicts a pressure drop between the ports that can be compared with the catheter measurements without any additional imaging."],"forward_implications":["If the closure and inference are sound, the same algorithm can be run with one- or two-equation turbulence models such as $k$–$\\varepsilon$; velocity data would constrain the RANS mean flow while turbulent-kinetic-energy data would constrain the turbulence model itself.","The assimilation produces a denoised, physically constrained mean velocity field at the model's finer resolution (0.75 mm) from 1 mm flow MRI data, effectively sharpening low-SNR scans without resolving turbulent eddies.","Quantified parameter uncertainties enable model comparison: different turbulence closures can be ranked by their marginal likelihood under the Laplace approximation rather than by ad hoc fit metrics.","Because any differentiable turbulence model can be inserted, the methodology extends to unsteady RANS, allowing time-resolved 4D flow MRI to be assimilated in the same framework.","For medical-device flows such as the FDA nozzle, this offers a path to device- or patient-specific closure parameters derived from routine MRI, reducing reliance on generic turbulence-model constants."],"supporting_citations":[{"why":"Supplies the adjoint-accelerated Bayesian inference machinery for reconstructing velocity fields from noisy flow MRI data and enforcing Navier–Stokes priors.","marker":"Kontogiannis et al. 2022"},{"why":"Defines the Bayesian inverse Navier–Stokes formulation, operator notation, and numerical stabilization parameters that the present RANS inversion inherits.","marker":"Kontogiannis et al. 2024a"},{"why":"Establishes simultaneous assimilation of velocity data and a non-uniform viscosity model, the direct predecessor extended here to turbulent eddy viscosity.","marker":"Kontogiannis et al. 2024b"},{"why":"Provides the Laplace approximation and Bayesian model-selection framework used to estimate posterior covariances around the MAP point.","marker":"MacKay 2003"},{"why":"Justifies modeling the flow MRI velocity noise as zero-mean Gaussian with the stated covariance, the data-likelihood assumption of the inverse problem.","marker":"Gudbjartsson & Patz 1995"},{"why":"Supplies the Bayesian inverse-problems formulation and the mathematical setting for Gaussian priors that grounds the problem statement.","marker":"Stuart 2010"}],"fun_headline_variants":["Bayesian inversion learns turbulence model from flow MRI","Flow MRI plus Bayesian stats yields turbulence model and uncertainties","Inverse RANS with MRI data learns eddy-viscosity parameters and errors","Learning turbulence closures from flow MRI: a Bayesian approach","MRI jet data teaches turbulence model via Bayesian inverse RANS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole inference rests on the assumption that a simple algebraic formula for the turbulence viscosity, built from a hand-chosen compound mixing length, can faithfully represent the turbulence of this jet in both the breakdown and pipe-flow regions, and this assumption is never checked against independent turbulence data.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian inversion learns turbulence model from flow MRI","Flow MRI plus Bayesian stats yields turbulence model and uncertainties","Inverse RANS with MRI data learns eddy-viscosity parameters and errors","Learning turbulence closures from flow MRI: a Bayesian approach","MRI jet data teaches turbulence model via Bayesian inverse RANS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000269,"raw_usage":{"total_tokens":1577,"prompt_tokens":854,"completion_tokens":723,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":641}},"tokens_in":470,"tokens_out":723,"duration_ms":6050,"temperature":1.0,"reasoning_tokens":641,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:06:55.335820+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the full turbulent stress tensor in the same FDA nozzle at Reynolds number 6500 with particle image velocimetry, and compare it with the stress predicted by the inferred eddy-viscosity parameters; if the two disagree beyond measurement error, the inferred parameters are artifacts of the closure rather than true properties of the turbulence.","supporting_citations":[{"cited_title":"V., Sederman, A","cited_arxiv_id":null,"evidence_quote":"Supplies the adjoint-accelerated Bayesian inference machinery for reconstructing velocity fields from noisy flow MRI data and enforcing Navier–Stokes priors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Laplace approximation and Bayesian model-selection framework used to estimate posterior covariances around the MAP point."},{"cited_title":"& Patz, S","cited_arxiv_id":null,"evidence_quote":"Justifies modeling the flow MRI velocity noise as zero-mean Gaussian with the stated covariance, the data-likelihood assumption of the inverse problem."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Bayesian inverse-problems formulation and the mathematical setting for Gaussian priors that grounds the problem statement."}],"review_version":1}