{"id":"207478f7-f2be-42b0-bb77-2e01eea1386c","arxiv_id":"2506.20341","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A DeepONet trained on LAMMPS platelet simulations reproduces time-resolved platelet deformation with sub-1% median error and extends with under 8% maximum error to held-out stiffness extremes.","lead":"The paper trains a neural-network model to reproduce, almost instantly, the full 3D deformation of a single platelet in shear flow, using data from slow particle-based simulations. It matters because a fast, accurate platelet model could one day make organ-scale, platelet-resolved blood flow simulations practical for thrombosis research.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Undefined 'relative displacement error' leaves the central sub-4% accuracy claim unverifiable; a rigid-body baseline should be run to check whether the reported errors are dominated by bulk translation rather than membrane deformation.","rationale":"The paper is a straightforward operator-learning study: a DeepONet maps (sigma, K, t, x0) to absolute particle positions, trained on LAMMPS DPD trajectories at one wall shear stress with ten bond stiffnesses. The extrapolation experiment in Sec. 4.1 is a genuine held-out test and should be credited. I agree with the reader's conditional assessment, but the single most load-bearing weakness is the undefined 'relative displacement error.' All central numbers (abstract, Sec. 5, Table 3, Figs. 4 and 8) are expressed in this metric, yet no formula or denominator is given. Since the model outputs absolute coordinates and the only dynamic validation shown is center-of-mass agreement, the reported error could be dominated by bulk translation. The Fig. 2 caption's attempted normalization is formally inconsistent (u_p(0) = 0), corroborating the ambiguity. A rigid-body baseline test would settle this: if it reproduces the reported error levels, the deformation claim collapses; if not, the metric still must be specified. Secondary issues noted by the reader—sigma fixed at 50 Pa despite being a network input, no code or data release, no leave-one-K cross-validation—are real limitations for deployment, but they do not undermine the calibrated-range claim as directly as the metric ambiguity does. The author's stated limitations (self-contact artifact, cost of leave-one-K retraining) are consistent with a conditional verdict. Minor internal inconsistencies, such as the 80% versus 90% train/validation split and the 10^8 sample cost estimate, should be corrected. I would not change the reader's verdict, but the acceptance conditions should explicitly require the metric definition and the rigid-body ablation.","tokens_in":8009,"tokens_out":10175,"duration_ms":110602,"concrete_test":"Run a rigid-body baseline on the same validation trajectories used for Table 3: at each time t, translate every particle by the LAMMPS ground-truth center-of-mass displacement (and apply the same rigid rotation, if any) while holding all coordinates relative to the center of mass fixed. Evaluate this baseline with the same 'relative displacement error' that produced Table 3 and Fig. 4. If the baseline's median and maximum errors are comparable to the reported 1% and 4%, the metric is dominated by bulk translation and the deformation-fidelity claim fails. If the baseline errors are much larger, the authors must still publish the exact metric definition and recompute errors on deformation-only variables, such as COM-subtracted coordinates or a strain measure, to substantiate the membrane-deformation claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is the undefined 'relative displacement error.' Every headline number—sub-1% median, <4% worst case, <8% extrapolation maximum, and all of Table 3—is expressed in this metric, yet no formula, denominator, or normalization is given anywhere in Secs. 3–4. The DeepONet outputs absolute particle coordinates x(t), and the only dynamic validation shown (Fig. 4, right) is the center-of-mass trajectory. In a Couette flow, bulk translation of the platelet over a Jeffery orbit can be much larger than the membrane's shape change, so a predictor that returns the true rigid-body motion while leaving the membrane shape fixed could score very low under a relative error normalized by total displacement. Such a predictor would not reproduce 'the full three-dimensional deformation of the membrane' claimed in Sec. 5. The Fig. 2 caption's attempted normalization, ||u(t)|| / max_p ||u_p(0)||, is formally inconsistent because u_p(0) = 0, which further shows the metric is not pinned down. Since every quantitative accuracy and extrapolation claim is stated only through this metric, the central claim is currently unverifiable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript trains a DeepONet surrogate to map (x0, t; sigma, K) to the absolute particle coordinates x(t) of a coarse-grained platelet membrane in a DPD Couette flow, with the capillary number varied by changing bond stiffness K at fixed wall shear stress. It reports median relative displacement errors below 1%, a 90th percentile below 3%, and a worst case below 4% across ten stiffness values, plus a leave-extremes-out extrapolation test with median error below 3% and maximum below 8%, and claims a four-to-five order-of-magnitude speed-up over LAMMPS. The paper concludes that the surrogate is suitable for coupling with continuum CFD in thrombosis simulations.","tokens_in":8228,"tokens_out":6136,"duration_ms":64322,"significance":"If the accuracy claims could be substantiated, the result would be a useful demonstration of operator learning for platelet mechanics and a concrete speed-up for multiscale blood-flow frameworks. The extrapolation design, which holds out the two extreme stiffnesses, is a genuine external check, and the per-capillary-number error table is internally consistent with the stated experiments. However, the central error metric is never defined, the per-capillary-number validation mixes training and held-out data, and the parameter sweep covers only one wall shear stress; until these are resolved, the headline numbers cannot be interpreted independently. The paper would benefit from code and data release and from a rigid-body baseline to separate bulk translation from membrane shape change.","major_comments":[{"comment":"The paper never defines the relative displacement error used in every headline result. The only normalization attempt appears in the Fig. 2 caption, ||u(t)|| / max_p ||u_p(0)||, but u_p(0) = 0 for all particles, so the denominator is zero. Because the network outputs absolute positions, a predictor that returns the correct rigid-body translation while leaving the membrane shape unchanged could score very differently depending on the chosen denominator. Please state the exact formula, including numerator, denominator, norm used, and whether center-of-mass or rigid rotation is subtracted, report shape-only errors separately from bulk translation errors, and include a rigid-body baseline for comparison.","section":"Sec. 3, Fig. 4, Table 3"},{"comment":"The per-capillary-number validation is not an independent test. The text says the network was trained on a random 90% subset of (x0, t, sigma, K) tuples, so 90% of the tuples at every K are in the training set, yet Table 3 is described as evaluating complete platelet trajectories, including those training points. Errors on memorized training data do not support the claim of a strict interpolation benchmark. Please report metrics only on the held-out 10% of tuples per K, or perform leave-one-K-out retraining; if full trajectories are used, separate train and test errors must be shown.","section":"Sec. 4, Table 3"},{"comment":"All data are generated at a single wall shear stress sigma = 50 Pa, with capillary number varied only by changing the bond constant K. Equation (4) nevertheless lists sigma as an input and the conclusions are framed as applying across physiologically relevant hemodynamic conditions. Varying K at fixed sigma changes the elastic response, but it is not shown to be equivalent to varying the flow-induced stress or shear rate. Please either add experiments at multiple sigma values, for example 50-250 Pa as in the prior work cited, or explicitly restrict the claims to the calibrated one-stress setting.","section":"Sec. 2.1, Table 2; Sec. 3, Eq. (4)"},{"comment":"The calibration from bond constant to capillary number is not specified. Equation (1) contains the in-plane shear modulus G_s, but the text only says G_s is linked to K without giving the conversion; Table 2 lists K ranges and resulting Ca* values without a formula or fitting procedure. Without this relationship, the reported Ca* values are not reproducible and the numerical values in Table 3 cannot be verified. Please provide the G_s(K) relation and any constants used.","section":"Sec. 2, Eq. (1), Table 2"}],"minor_comments":[{"comment":"The dataset size is stated inconsistently: Sec. 2.2 says roughly 18,000 particles and 101 time instants, giving about 1.8 x 10^6 states per trajectory, while Sec. 4 says a leave-one-K-out retraining would involve data sets exceeding 10^8 samples. Please reconcile these numbers.","section":"Sec. 2.2, Sec. 4"},{"comment":"The speed-up claim of four to five orders of magnitude is not supported by any timing measurement or code; please add wall-clock benchmarks for both LAMMPS and the DeepONet inference.","section":"Sec. 5"},{"comment":"Typo: 'most complaint platelet' should be 'most compliant platelet'.","section":"Fig. 6 caption"},{"comment":"The phrase 'dimensional capillary number' is misleading because Eq. (1) is dimensionless; consider removing 'dimensional'.","section":"Sec. 2.1"},{"comment":"No data or code availability statement is included; providing the LAMMPS input scripts, trained weights, and evaluation script would substantially strengthen reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper relies heavily on two prior self-cited preprints for baseline claims, and no code or data link is provided; the editor may wish to require an external reproducibility check. The central error metric should be defined before any acceptance decision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuine step beyond the author's earlier DeepONet platelet papers. Instead of end-state shapes or a single shear stress, the surrogate learns full time-resolved trajectories across ten capillary numbers, and the leave-extremes-out experiment is a real held-out test. The error tables are internally consistent, and the error peaks at membrane self-contact are a plausible, well-identified failure mode. So the paper earns a serious look.\n\nThe soft spots are real. The biggest one is exactly what the stress-test note flags: the 'relative displacement error' is never defined. Every headline number, including the sub-1% median and the <4% worst case, is expressed in that metric, but there is no formula anywhere in Sections 3–4. The Fig. 2 caption's attempted normalization is self-inconsistent because u_p(0) = 0, so the denominator is zero. That means the central accuracy claim is currently unverifiable. It is entirely possible that the errors are dominated by bulk translation of the platelet in the Couette flow rather than by membrane shape change, and a rigid-body baseline would settle that. The paper shows center-of-mass agreement, but that does not demonstrate membrane fidelity.\n\nSecond concern: the capillary-number sweep is done by changing the spring constant at a fixed shear stress of 50 Pa. So the 'across Ca*' claim is really a claim about stiffness at one flow condition. Whether that transfers to hemodynamic settings with different shear stresses is not established. The network does take sigma as an input, but only one value is used, so that input is effectively untested.\n\nThird, no code or data are provided, which matters for a benchmark paper. And the interpolation test is a random 90% split of tuples, not a hold-out of entire K levels; every K appears in training. The extrapolation study is the genuinely clean test, and it does support graceful degradation.\n\nNone of this is fatal. The author explicitly acknowledges the leave-one-K cost and the self-contact artifact, and the extrapolation result is real evidence. But until the metric is defined and a translation-only baseline is run, I would take the sub-1% claim as provisional.\n\nRecommendation: send it to peer review. It is a legitimate technical contribution that a good referee can push into shape. Require the metric definition, the rigid-body baseline, and preferably code/data before acceptance.","headline":"A real extension of the prior DeepONet platelet work, but the central accuracy metric is undefined and the Ca* sweep is only a stiffness sweep at one shear stress.","tokens_in":8747,"tokens_out":2060,"would_cite":false,"duration_ms":21879,"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 DeepONet trained on platelet simulations spanning capillary numbers 0.07–0.77 yields full 3D deformation trajectories with sub-percent median error, worst cases under 4%, and a 10,000–100,000x speed-up.","keywords":["neural operator","DeepONet","platelet deformation","capillary number","dissipative particle dynamics","LAMMPS","multiscale thrombosis modeling","shear flow"],"falsifier":"Recompute the relative displacement error after subtracting the platelet's center-of-mass trajectory from both the LAMMPS ground truth and the DeepONet output: if the residual is materially above the reported sub-percent median, the headline metric conflates bulk motion with deformation. A complementary test is to hold the stiffness fixed and vary the imposed shear rate or stress instead, since the training data never move $\\sigma$ off 50 Pa; if accuracy degrades there, the claim of fidelity 'across capillary numbers' does not transfer to other flow conditions.","tokens_in":7773,"feed_emoji":"🩸","tokens_out":19609,"duration_ms":173015,"temperature":0.7,"pith_summary":"This paper aims to establish that a DeepONet neural operator, trained on LAMMPS particle simulations of a spring-network platelet membrane, can stand in for the simulation across the physiologically relevant capillary-number range. Ten bond-stiffness settings produce capillary numbers from 0.07 to 0.77, and the network maps initial particle coordinates, time, wall shear stress, and stiffness to the instantaneous three-dimensional position of every membrane particle. The authors report a median relative displacement error below 1%, a 90th percentile below 3%, and a worst case below 4%, with single-evaluation speed-ups of four to five orders of magnitude. If these numbers hold, multiscale thrombosis models could run platelet-resolved hemodynamics in patient-specific geometries at a cost that pure particle simulation cannot reach.","feed_headline":"Sub-1% error at 10,000x speed: AI surrogate for platelet deformation","feed_subtitle":"One network call replaces a costly particle simulation—a step toward organ-scale thrombosis models.","key_machinery":"The load-bearing object is the DeepONet operator: a branch fully connected network (two hidden layers, 32 and 16 nodes) that encodes the scalar inputs $(\\sigma, K)$, and a trunk fully connected network (three hidden layers, 32, 32, and 16 nodes) that encodes $(x_0, t)$, joined by an inner product that outputs the particle's absolute position $x(t)$. The training data come from LAMMPS simulations in which dissipative particle dynamics drives a hollow ellipsoidal platelet built from roughly 18,000–20,000 particles linked by harmonic bonds of stiffness $K$; the capillary number $\\mathrm{Ca}^* = \\mu\\dot{\\gamma}a/G_s$ is varied by sweeping $K$ while the wall shear stress stays at 50 Pa. What carries the argument is the per-capillary-number error audit: the network is trained on a random 90% of state tuples, so every stiffness appears in training, but it is validated on complete trajectories that include time instants and particle indices never seen. The paper reports the median, 90th-percentile, and maximum relative displacement error for each of the ten stiffnesses, and repeats the protocol with the two extreme cases held out entirely to support the extrapolation claim.","core_discovery":"The central claim is that the operator $G: (x_0, t; \\sigma, K) \\mapsto x(t)$ — the map from a membrane particle's initial position, the elapsed time, the wall shear stress, and the membrane bond stiffness to that particle's current coordinates — is learnable to high fidelity by a fully connected DeepONet with a branch network for $(\\sigma, K)$ and a trunk network for $(x_0, t)$. Trained on roughly $10^6$ particle states per trajectory over ten capillary numbers ($0.07 \\le \\mathrm{Ca}^* \\le 0.77$), the network keeps the median relative displacement error under 1%, the 90th percentile under 3%, and the maximum under 4% across the calibrated range, while accelerating the particle simulation by four to five orders of magnitude. With the two extreme stiffnesses left out of training, the median error stays below 3% and the maximum near 8%, which the authors read as graceful extrapolation. Error analysis places the largest deviations at membrane self-contact events and in high-curvature regions; the authors attribute these to the kinematic discontinuity of the hollow-membrane idealization and propose graph-neural trunks and physics-informed torque regularization as remedies. The larger claim is that a surrogate in this accuracy class is fit for coupling with continuum CFD in platelet-resolved thrombosis simulations.","pith_inferences":["The capillary-number sweep is realized only through bond stiffness at one fixed shear stress of 50 Pa, even though $\\sigma$ is a network input; a natural stress test the paper does not run is varying the shear rate or stress with stiffness fixed, since real vessels expose platelets to a broad stress range.","The headline metric deserves a bulk-motion check: recomputing the error after subtracting the platelet's center-of-mass translation from both prediction and ground truth would show whether the sub-percent figure describes shape change or mainly advection.","The self-contact error peaks are artefacts of the hollow-shell idealization; a filled or cytoskeleton-bearing platelet model would remove them, so the surrogate's stated worst-case errors may largely shrink in more realistic geometries.","The speed-up claim refers to a single micro-solver evaluation; the net gain inside a coupled CFD loop also depends on how many surrogate calls per cell per time step the coupling requires, which the paper does not quantify."],"forward_implications":["One DeepONet evaluation replaces an individual LAMMPS micro-simulation in a multiscale loop, cutting the micro-solver cost by four to five orders of magnitude.","Within the calibrated capillary-number range the surrogate stays under 4% worst-case error, which the paper classifies as high-fidelity enough to couple with continuum CFD for platelet-resolved hemodynamics.","The leave-extremes-out retraining keeps sub-3% median error on the stiffest and most compliant platelets, so the usable envelope extends beyond the ten trained stiffnesses.","The error audit identifies membrane self-contact and high-curvature patches as the dominant failure modes, directing future work toward graph-neural trunks and physics-informed torque regularization.","Because the surrogate resolves full time histories (101 instants per trajectory) rather than end states, it can feed time-resolved deformation data into shear-driven models of platelet transport, margination, and adhesion."],"supporting_citations":[{"why":"Supplies the DeepONet architecture — branch and trunk fully connected networks joined by an inner product — that the surrogate is built on.","marker":"[2]"},{"why":"The authors' earlier workflow, sampling 101 shear stresses but recording only end-state platelet shapes, which this paper extends to full trajectories.","marker":"[3]"},{"why":"The prior single-shear-stress (50 Pa) full-time-evolution study whose dataset and training design this paper expands to ten capillary numbers.","marker":"[4]"},{"why":"Defines the capillary number as the ratio of viscous to elastic forces and motivates its role in platelet deformation and margination, grounding the parameter range studied.","marker":"[20]"},{"why":"The LAMMPS particle-simulation code that generates every training trajectory the surrogate learns from.","marker":"[27]"},{"why":"Supplies the dissipative particle dynamics method used to model the surrounding blood flow.","marker":"[28]"},{"why":"Provides the multiscale particle-based platelet-in-plasma model (DPD fluid plus coarse-grained membrane) that the spring-network platelet representation follows.","marker":"[29]"}],"fun_headline_variants":["AI surrogate for platelet deformation hits sub-1% median error","Neural operator accelerates platelet deformation simulation by 10,000x","Platelet deformation surrogate: sub-1% median error, 10,000x speedup","DeepONet predicts platelet shape to sub-1% error with 10,000x speedup"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracy presumes that varying the membrane's spring stiffness at a fixed wall shear stress of 50 Pa exposes the same deformation regimes as varying the flow itself, and that the paper's 'relative displacement error' measures membrane shape change rather than the platelet's bulk translation.","fun_headline_variants_meta":{"raw":{"variants":["AI surrogate for platelet deformation hits sub-1% median error","Neural operator accelerates platelet deformation simulation by 10,000x","Platelet deformation surrogate: sub-1% median error, 10,000x speedup","DeepONet predicts platelet shape to sub-1% error with 10,000x speedup"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000842,"raw_usage":{"total_tokens":3718,"prompt_tokens":1043,"completion_tokens":2675,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":659,"completion_tokens_details":{"reasoning_tokens":2588}},"tokens_in":659,"tokens_out":2675,"duration_ms":17104,"temperature":1.0,"reasoning_tokens":2588,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:51:16.655047+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the relative displacement error after subtracting the platelet's center-of-mass trajectory from both the LAMMPS ground truth and the DeepONet output: if the residual is materially above the reported sub-percent median, the headline metric conflates bulk motion with deformation. A complementary test is to hold the stiffness fixed and vary the imposed shear rate or stress instead, since the training data never move $\\sigma$ off 50 Pa; if accuracy degrades there, the claim of fidelity 'across capillary numbers' does not transfer to other flow conditions.","supporting_citations":[{"cited_title":"Learning nonlinear operators via deeponet based on the uni- versal approximation theorem of operators.Nature machine intelligence, 3(3):218–229, 2021","cited_arxiv_id":null,"evidence_quote":"Supplies the DeepONet architecture — branch and trunk fully connected networks joined by an inner product — that the surrogate is built on."},{"cited_title":"High-Fidelity Description of Platelet Deformation Using a Neural Operator","cited_arxiv_id":"2412.00747","evidence_quote":"The authors' earlier workflow, sampling 101 shear stresses but recording only end-state platelet shapes, which this paper extends to full trajectories."},{"cited_title":"Neural Operator Modeling of Platelet Geometry and Stress in Shear Flow","cited_arxiv_id":"2503.12074","evidence_quote":"The prior single-shear-stress (50 Pa) full-time-evolution study whose dataset and training design this paper expands to ten capillary numbers."},{"cited_title":"Effect of tube diameter and capillary number on platelet margination and near-wall dynamics.Rheologica Acta, 55:511–526, 2016","cited_arxiv_id":null,"evidence_quote":"Defines the capillary number as the ratio of viscous to elastic forces and motivates its role in platelet deformation and margination, grounding the parameter range studied."},{"cited_title":"Dissipative particle dynamics: Bridging the gap between atomistic and mesoscopic simulation.The Jour- nal of chemical physics, 107(11):4423–4435, 1997","cited_arxiv_id":null,"evidence_quote":"Supplies the dissipative particle dynamics method used to model the surrounding blood flow."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the multiscale particle-based platelet-in-plasma model (DPD fluid plus coarse-grained membrane) that the spring-network platelet representation follows."}],"review_version":1}