{"id":"75fae001-fcfd-4c8f-8f4c-79b61200ab91","arxiv_id":"2412.14572","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A differentiable 1D hemodynamics solver in JAX enables gradient-based inference of Windkessel parameters from pressure waveforms.","lead":"This paper presents a blood-flow simulator for arteries that can compute gradients of its predictions with respect to patient-specific settings. The authors use those gradients to calibrate the model to synthetic patient data quickly, with the goal of making personalized cardiovascular modeling faster and more interpretable.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Optimizing on unconverged transients is the load-bearing unsupported step: the reported calibration success is not shown to transfer to the periodic steady state, and no gradient-free baseline supports the 'accelerated' claim.","rationale":"The forward-solver validation against openBF (Table 1, Figure 1) is the strongest part of the paper and is credible; I do not contest the differentiability or the basic solver construction. The central claim, however, is about patient-specific calibration speed, and that claim fails where the calibration evidence is weakest. The paper's own text in Section 3.5.1 concedes that the simulations were not converged. A patient-specific calibration target is the periodic steady-state pressure waveform; if the optimizer is fitting a transient window, the recovered RCR parameters may be an artifact of the initial transient. The two synthetic demonstrations are reassuring but not sufficient: they use data generated by the same solver, start relatively close to the truth, and provide no converged-state comparison. The negative loss values reported in Tables 5-7, for a loss defined as a positive squared-error sum, also make the quantitative convergence evidence difficult to interpret. A simple computational test, re-running the same optimizations after the forward solver has reached the paper's own convergence tolerance, would directly settle whether the unconverged-gradient shortcut is valid. The reader's CONDITIONAL verdict is therefore appropriate; this stress-test does not change it.","tokens_in":29519,"tokens_out":7311,"duration_ms":56747,"concrete_test":"Re-run the two deterministic inference cases in Section 3.5.1 with identical geometry, ground truth, initialization, and Adafactor settings, but first evolve each forward simulation to periodic convergence using the Appendix C criterion (pressure change < 0.1 mmHg between cardiac cycles); compare the recovered R1/R2 values and final loss to Tables 5-6. If the converged-calibration parameters differ materially or require many more iterations, the unconverged-transient shortcut is the unsupported load-bearing assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that gradients from a short, unconverged simulation window correctly steer Windkessel parameters toward the values that reproduce the patient's periodic pressure waveform. Section 3.5.1 explicitly says the simulation was run for time steps 'despite our simulation not having converged yet'; the deterministic and probabilistic inference experiments use N=1000 (or N=1) steps in this transient. The two synthetic examples recover the known RCR values, but this shows only that those particular transients are informative. Nothing in the manuscript establishes, analytically or empirically, that the transient loss landscape is aligned with the converged periodic landscape. If they are not aligned, calibration can match the truncated simulation while giving wrong R1/R2 for the actual periodic patient state. The absence of any comparison to gradient-free calibration methods means the abstract's 'faster' claim is supported only by forward-solve wall-clock timings (Figure 2) and not by inverse-problem cost measurements.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents jaxFlowSim, a differentiable 0D-1D reduced-order hemodynamics solver implemented in JAX using a MUSCL finite-volume scheme, Windkessel outlet models, and junction treatments. The forward solver is validated against the open-source openBF solver on four vascular anatomies, with relative L1 pressure errors below 1%. Wall-clock timings show speedups over openBF for anatomies with more than roughly 15 vessel segments. The authors then use gradients through the solver to perform deterministic (Adafactor) and probabilistic (Hamiltonian Monte Carlo) inference of Windkessel RCR parameters on a single-bifurcation synthetic geometry, reporting convergence to known ground-truth values within one to four minutes.","tokens_in":29724,"tokens_out":6853,"duration_ms":39832,"significance":"If the parameter-inference claims are substantiated, the differentiable solver would be a genuinely useful tool for interpretable, gradient-based calibration of reduced-order hemodynamic models. The forward validation against an independent solver is a concrete strength, as is the engineering achievement of near-constant compile time and roughly linear compute scaling in the number of vessel segments. The current inference experiments, however, are synthetic self-consistency checks on truncated transients with no comparison to gradient-free baselines, so the central 'accelerated patient-specific calibration' claim is not yet established by the evidence presented.","major_comments":[{"comment":"The loss function L(ξ) defined in §3.5.1 is a normalized sum of squared errors and is therefore nonnegative, yet all loss values in Tables 5, 6, and 7 are negative (e.g., −5.77, −10.25, −13.41). This inconsistency makes the reported convergence behavior uninterpretable from the stated objective. Please either correct the reported values or clarify that a different objective, such as a negative log-likelihood, is being tabulated, and define that objective explicitly.","section":"§3.5.1, Tables 5–7"},{"comment":"The optimization is deliberately run on non-converged, truncated simulations: the text states 'we simulated for time-steps despite our simulation not having converged yet.' The paper provides no argument or experiment showing that gradients computed from this transient window point toward the parameters that reproduce the converged periodic pressure waveform. Because patient calibration ultimately targets the periodic steady state, the recovered R1 and R2 values could match the truncated transient while being incorrect for the actual periodic state. Please demonstrate, for the same test cases, that the parameters obtained from the truncated loss also reproduce the converged periodic solution, e.g., by re-running the forward solver with the optimized parameters until convergence and reporting the periodic-state pressure error.","section":"§3.5.1, paragraph before Fig. 9"},{"comment":"The abstract and introduction claim 'faster' and 'accelerated' parameter inference, but no comparison against gradient-free calibration methods is provided for the inverse problem. The wall-clock times in Tables 5–7 are absolute and are not contrasted with, for example, Nelder–Mead, CMA-ES, or Bayesian optimization on the same calibration task. Figure 2 compares only forward-solve timings. To support the acceleration claim, please add inverse-problem wall-clock times and iteration counts for at least one standard gradient-free baseline.","section":"§3.5, Tables 5–7"},{"comment":"The HMC section states 'We set N = 1000, implying that only the last time step that we simulated is used in our loss function.' This contradicts the earlier usage of N as the number of simulated time steps and the statement in §3.5.1 that N = 1000 means all simulated time steps are used in the loss. Please clarify which time steps enter the HMC likelihood and correct the inconsistency; the current text leaves the exact likelihood ambiguous and hampers reproducibility of the posterior results.","section":"§3.5.2, paragraph after Fig. 10"}],"minor_comments":[{"comment":"The text says measurements were evaluated at M = 2 locations 'at the end of each outlet vessel,' but Figure 9 is described as showing pressure in the middle of the first outlet vessel. Please clarify the measurement locations and match them to the plotted waveforms.","section":"§3.5.1, first case"},{"comment":"The initialization is described as ξ0 = {2 × log(1 + e^{s0}) × ξ̂}, and the table column is labeled 'scaling factors' but lists values that equal softplus(s0), not s0 itself. Please define and label the tabulated quantity consistently.","section":"§3.5.1, initialization"},{"comment":"The density ρ is written as a function of z and t in Eq. (10), although the reduction assumed blood density to be constant. Please use a constant ρ or explain the retained dependence.","section":"§2.2, Eq. (10)"},{"comment":"The parallel-circuit distribution uses ξ_i = Σ_j A_{0,j}/A_{0,i}, but the derivation is not shown. Please provide a brief explanation of why resistance is distributed inversely with reference cross-sectional area and how the parallel-circuit rule is applied.","section":"§3.1, Windkessel parameter distribution"},{"comment":"The text calls the 19-vessel geometry an outlier in the compute-time scaling, but no quantitative justification is given. Please either report the timings for all four geometries or explain the source of the apparent outlier.","section":"§3.4, Figure 7"}],"recommendation":"major_revision","confidential_remarks":"The forward-solver validation is solid and the differentiable implementation is a useful contribution. The main gap is that the central 'accelerated calibration' claim rests on synthetic, unconverged-transient experiments with no gradient-free baseline and with negative loss values that contradict the stated loss function. These issues are fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know one thing up front: the forward solver is the real contribution, and it is solid. jaxFlowSim is a JAX implementation of a standard 1D-0D hemodynamics model (MUSCL-FV, Windkessel outlets, junction conditions). The authors validate it against openBF on four anatomies (9 to 77 vessels) with relative L1 pressure errors below 1%, and show near-linear wall-clock scaling and roughly 2x speedup on the largest case. That is genuine, useful engineering, and it is the part of the paper I trust.\n\nThe parameter inference sections are where I get off the train. The 'accelerated patient-specific calibration' headline is supported only by synthetic-data experiments: ground truth generated by the same solver, same initial conditions, and the optimizer recovers those known parameters. That is self-consistency, not calibration. There is no comparison to a gradient-free or standard calibration method, so the 'faster' claim in the abstract is not measured. The text also admits in Section 3.5.1 that the optimization was run on a simulation that had not converged to periodic steady state. If the clinical target is the patient's periodic pressure waveform, nothing in the paper shows that gradients from a truncated transient window point at the right steady-state parameters. This is a genuine unresolved issue, not a nitpick. It is possible the transient is informative enough in practice, but the paper needs either to run to convergence or provide evidence that the transient loss landscape aligns with the converged one.\n\nThere are also smaller presentation problems that a referee should catch: the loss function is written with undefined indices, and the tables report negative loss values for what should be a sum of squared errors. No code or data are released, which matters for a tool paper.\n\nNet: this is a worthwhile paper for hemodynamics modellers who want a differentiable 1D solver, and the forward validation deserves credit. The inference experiments are a proof of concept, not a demonstration of accelerated patient-specific calibration. I would send it to peer review because the solver itself is useful and the forward validation is credible, but I would not accept it in this form; the calibration claims need either to be scaled back or backed by convergence-aware experiments and a baseline comparison.","headline":"A genuinely useful differentiable 1D hemodynamics solver with solid forward validation, but the calibration claims rest on synthetic self-consistency and an unexamined transient-optimization shortcut.","tokens_in":30184,"tokens_out":2922,"would_cite":true,"duration_ms":23843,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M08","76Z05","92C35"],"pacs":["47.63.Cb","87.19.Uv"],"model":"deepseek-v4-flash","headline":"A differentiable 1D-0D hemodynamics solver enables gradient-based inference of patient-specific vessel and Windkessel parameters from pressure data, with forward predictions matching an established solver within 1% error.","keywords":["differentiable hemodynamics","1D Navier-Stokes reduced-order model","patient-specific calibration","Windkessel parameter inference","JAX automatic differentiation","MUSCL finite volume scheme","Hamiltonian Monte Carlo"],"falsifier":"Repeat the four-parameter resistance inference on the same bifurcation geometry, but this time iterate the solver to the paper's own convergence criterion (pressure changing by less than 0.1 mmHg between cardiac cycles) before taking gradients, and compare the recovered $R_1$ and $R_2$ values with those obtained from the truncated unconverged window; if any parameter differs by more than the optimizer's tolerance, the accelerated protocol is fitting a transient that the converged periodic state would contradict.","tokens_in":29274,"feed_emoji":"🫀","tokens_out":12602,"duration_ms":98789,"temperature":0.7,"pith_summary":"This paper tries to establish that a reduced-order (0D-1D) model of blood flow can be made fully differentiable without sacrificing the interpretability of the underlying physics, and that this differentiability turns patient-specific calibration into fast optimization. The solver, written in the JAX library, computes pressure waveforms through a finite-volume discretization of the 1D Navier-Stokes equations and at the same time returns exact gradients of those waveforms with respect to vessel lengths, stiffnesses, resistances, and compliances. With those gradients, the paper recovers Windkessel parameters from synthetic pressure measurements by gradient descent in about a minute and by Hamiltonian Monte Carlo in a few minutes, and it validates the solver's forward outputs against the open-source openBF code on four anatomies, finding less than 1% relative error and a speedup that approaches two times on the largest network. If correct, this gives a middle path between slow gradient-free calibration and uninterpretable deep-learning surrogates: fast, physics-grounded inference of parameters that cannot be measured noninvasively, with sensitivity analysis available from the same gradients.","feed_headline":"Differentiable blood-flow solver calibrates patient models in minutes","feed_subtitle":"Matches an established solver within 1% error while unlocking fast, interpretable parameter inference.","key_machinery":"The load-bearing object is the differentiable solver itself, which is a finite-volume discretization of the dimension-reduced Navier-Stokes equations for a compliant tube: conservation laws for cross-sectional area $A$ and flow rate $Q$, closed by the tube-law pressure relation $P = P_{\\mathrm{ext}} + \\beta(\\sqrt{A/A_0}-1)$ with elasticity coefficient $\\beta$. A MUSCL scheme with total-variation-diminishing slope limiting and Lax-Friedrichs fluxes advances the hyperbolic system; inlet values come from Riemann-invariant extrapolation of given pressure or flow data, and outlets are coupled to three-element Windkessel (RCR) models whose roots are found by a frozen Newton method. Because the whole chain is implemented in JAX with padding and masking to vectorize across vessels, automatic differentiation yields gradients of any simulated quantity with respect to any parameter, and it is these gradients that both the Adafactor and the Hamiltonian Monte Carlo inference procedures consume.","core_discovery":"Stated on the paper's own terms, the discovery is that the full pipeline — cross-sectionally averaged 1D Navier-Stokes equations, tube-law pressure closure, MUSCL finite-volume time stepping, Riemann-invariant boundary conditions, and Windkessel outlets solved by Newton's method — can be written in JAX so that automatic differentiation flows through every step. The paper demonstrates two consequences of this differentiability. First, calibration: minimizing a normalized pressure error with the Adafactor optimizer recovers a single terminal resistance close to its ground-truth value within about 500 iterations (under a minute), and recovers four Windkessel resistances of a two-outlet bifurcation within 1000 iterations (under two minutes), while Hamiltonian Monte Carlo yields a posterior concentrated at the same values after 500 sampling steps in less than four minutes. Second, fidelity: pressure waveforms from jaxFlowSim overlap those of openBF on aortic, abdominal, cerebral, and full-body ADAN56 anatomies with relative errors between $2.5 \\times 10^{-4}$ and $5.5 \\times 10^{-3}$, with the JAX implementation nearly twice as fast on the 77-segment network.","pith_inferences":["The differentiability the paper builds for parameter calibration also invites learning the tube law itself from data: the authors note that the pressure–area relation is empirical and contested, and the solver-in-a-loop idea they sketch would let pressure measurements select among or refine such constitutive laws.","The authors deliberately optimize on simulations that have not reached periodic steady state; a natural extension is to test how few cardiac cycles suffice for reliable gradients, since each skipped cycle multiplies the calibration speedup.","The junction solves, flagged as not readily GPU-parallelizable, are the likely scaling bottleneck; reformulating all junctions as one sparse system is the concrete step that would let GPU batching of many patients pay off.","Gradients of pressure with respect to measurement location could guide experimental design: placing sensors where gradient norms are largest would yield the most informative data for identifying RCR parameters."],"forward_implications":["Windkessel resistances and compliances — parameters that cannot be measured noninvasively — can be recovered from pressure data by gradient descent in roughly a minute for a single parameter and under two minutes for four parameters.","Hamiltonian Monte Carlo inference over four parameters is close to ground truth after 500 sampling steps, completing in less than four minutes, which makes posterior uncertainty estimates practical in a calibration setting.","The solver's forward predictions agree with openBF to within 1% relative error on networks from 9 to 77 vessels, and wall-clock time scales almost linearly with vessel count while compile time stays near constant.","Sensitivity analysis of any simulated biomarker with respect to vessel and outlet parameters becomes a single backward pass, which the paper positions as a step toward deciding which patient-specific measurements are informative.","JAX's device-agnostic execution means the same solver can run batched on a GPU, so many patient models could be calibrated in parallel."],"supporting_citations":[{"why":"JAX library: supplies the automatic differentiation through the solver that the entire calibration claim rests on.","marker":"[30]"},{"why":"openBF: the established 1D hemodynamics solver whose pressure waveforms serve as the validation baseline for jaxFlowSim.","marker":"[82]"},{"why":"Blanco et al.: provides the ADAN56 network with its carefully calibrated RCR parameters, plus the empirical wall-thickness and stiffness relations used to configure anatomies.","marker":"[83]"},{"why":"Vascular Model Repository: source of the three-dimensional patient anatomies converted into solver configurations.","marker":"[81]"},{"why":"Van Leer: the MUSCL finite volume scheme with total-variation-diminishing limiting that discretizes the 1D equations.","marker":"[68]"},{"why":"Formaggia et al.: the tube-law pressure–area closure and the outlet modeling conventions the solver adopts.","marker":"[47]"},{"why":"Parker et al.: the three-element Windkessel (RCR) outflow model whose resistances and compliance are the inference targets.","marker":"[74]"},{"why":"Shazeer and Stern: the Adafactor optimizer used for the deterministic gradient-based parameter recovery.","marker":"[90]"},{"why":"Phan et al.: NumPyro, the Hamiltonian Monte Carlo engine used for the probabilistic parameter inference.","marker":"[92]"}],"fun_headline_variants":["Differentiable blood-flow solver nails patient calibration in minutes","Fast patient-specific calibration via differentiable hemodynamics","Gradient-based blood flow calibration: fast, interpretable, accurate","Match openBF accuracy with JAX differentiable solver in minutes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The optimization experiments stop the solver before the flow has settled into its repeating heartbeat cycle, and the claim of fast calibration assumes that gradients computed from that truncated, not-yet-converged window still point toward the parameters of the true periodic state.","fun_headline_variants_meta":{"raw":{"variants":["Differentiable blood-flow solver nails patient calibration in minutes","Fast patient-specific calibration via differentiable hemodynamics","Gradient-based blood flow calibration: fast, interpretable, accurate","Match openBF accuracy with JAX differentiable solver in minutes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00098,"raw_usage":{"total_tokens":4217,"prompt_tokens":1060,"completion_tokens":3157,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":676,"completion_tokens_details":{"reasoning_tokens":3092}},"tokens_in":676,"tokens_out":3157,"duration_ms":21902,"temperature":1.0,"reasoning_tokens":3092,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:06:37.909510+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the four-parameter resistance inference on the same bifurcation geometry, but this time iterate the solver to the paper's own convergence criterion (pressure changing by less than 0.1 mmHg between cardiac cycles) before taking gradients, and compare the recovered $R_1$ and $R_2$ values with those obtained from the truncated unconverged window; if any parameter differs by more than the optimizer's tolerance, the accelerated protocol is fitting a transient that the converged periodic state would contradict.","supporting_citations":[{"cited_title":"JAX: composable transfor- mations of Python+NumPy programs, 2018","cited_arxiv_id":null,"evidence_quote":"JAX library: supplies the automatic differentiation through the solver that the entire calibration claim rests on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"openBF: the established 1D hemodynamics solver whose pressure waveforms serve as the validation baseline for jaxFlowSim."},{"cited_title":"An anatomically detailed arterial network model for one-dime nsional computational hemodynamics","cited_arxiv_id":null,"evidence_quote":"Blanco et al.: provides the ADAN56 network with its carefully calibrated RCR parameters, plus the empirical wall-thickness and stiffness relations used to configure anatomies."},{"cited_title":"V ascular model repository, 2023","cited_arxiv_id":null,"evidence_quote":"Vascular Model Repository: source of the three-dimensional patient anatomies converted into solver configurations."},{"cited_title":"Towards the ultimate conservative diffe rence scheme","cited_arxiv_id":null,"evidence_quote":"Van Leer: the MUSCL finite volume scheme with total-variation-diminishing limiting that discretizes the 1D equations."},{"cited_title":"Cardiovascular Mathematics: Modeling and simulation of the circulatory system , volume 1","cited_arxiv_id":null,"evidence_quote":"Formaggia et al.: the tube-law pressure–area closure and the outlet modeling conventions the solver adopts."},{"cited_title":"Lumped parameter out flow models for 1-d blood flow simulations: Effect on pulse waves and parameter estimation","cited_arxiv_id":null,"evidence_quote":"Parker et al.: the three-element Windkessel (RCR) outflow model whose resistances and compliance are the inference targets."},{"cited_title":"Adafactor: Adaptive l earning rates with sublinear memory cost","cited_arxiv_id":null,"evidence_quote":"Shazeer and Stern: the Adafactor optimizer used for the deterministic gradient-based parameter recovery."}],"review_version":1}