{"id":"db2d6c0a-2ebd-4ad1-bc96-c6cdb751e9d9","arxiv_id":"2412.00747","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A DeepONet neural operator reproduces LAMMPS platelet deformation simulations under shear flow with a typical error of about 0.5%.","lead":"This paper tests whether a neural network called DeepONet can quickly predict how a blood platelet deforms under shear flow, using data from particle simulations. A generalist reader might care because such fast surrogates could one day link microscopic platelet behavior into large-scale blood flow models of clotting.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Concern: The 0.5%/2% accuracy claim is measured against simulation output, not the underlying model's fidelity, and the reported error metric normalizes by the full displacement vector, which may obscure particle-wise errors.","rationale":"The reader's stated weakest assumption is the faithfulness of the DPD model to real platelets, which is a relevant external-validity concern, but the paper itself claims only to surrogate the simulation, so that concern, while legitimate, does not directly undermine the internal claim of 'captures the dynamics of the membrane particles.' The more load-bearing internal concern is that the paper's quantitative claim of '0.5% mode error' is presented as a measure of fidelity of the deformation description, but Eq. (7) defines an error about the full displacement vector. In a Couette-flow Jeffery orbit, the displacement is dominated by rigid-body rotation/translation of the platelet, not by membrane strain. Thus, the reported low error may be trivially achievable by a model that predicts only the bulk rotation and misses the biologically relevant membrane deformation. The paper does not provide any test that separates rigid-body motion from deformation, so the central claim of 'high-fidelity description of platelet deformation' is not yet supported by the evidence presented. The concern is testable by an alignment-based analysis. I therefore concur with the CONDITIONAL verdict but for a more precise reason than the reader's stated weakest assumption. The reader partially identified the issue of generalization to different initial configurations and the test-set architecture selection, but did not flag the rigid-body-motion contamination of the error metric, which I consider the single most load-bearing internal issue. The external DPD-fidelity concern is secondary because the paper's stated goal is to surrogate the simulation. A concrete test is to align the predicted and true configurations and recompute errors on the residual deformation field; if the aligned error is also low, the claim survives. If not, the headline claim should be weakened to 'low displacement error' rather than 'high-fidelity deformation.' Hence the verdict should remain CONDITIONAL, with the condition being demonstration of deformation-only accuracy and, ideally, release of code/data for reproducibility. I do not see a reason to reject, as the reported interpolation is internally plausible and the disclosed limitations are honest. No ad hominem is intended; the critique is solely about the mismatch between the reported metric and the claim it is used to support.","tokens_in":9963,"tokens_out":2203,"duration_ms":18471,"concrete_test":"Recompute the relative difference using a deformation-only metric, e.g., remove the rigid-body translation and rotation from both the predicted and ground-truth configurations (using a best-fit rotation in the least-squares sense, such as a Kabsch-like alignment) and then compute the per-particle relative difference in the residual deformation field. Also compute the strain tensor or the radial/azimuthal displacement of membrane particles relative to the fitted ellipsoid. If the aligned deformation error is significantly larger than 2%, the headline accuracy claim overstates fidelity. In addition, report the fraction of the displacement norm attributable to rigid-body rotation versus deformation.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim, as stated in the abstract and Section 3.2, is that the neural operator captures platelet deformation dynamics with a mode error distribution of approximately 0.5% and a maximum relative difference below 2%. This claim is load-bearing because the entire contribution is framed as high-fidelity surrogate modeling for thrombosis. The concern is not that the DeepONet fails to interpolate the LAMMPS DPD data, but that the reported error metric is potentially misleading and under-tested. The relative difference defined in Eq. (7), D = ||xi_pred - xi_true|| / ||xi_true||, normalizes by the norm of the ground truth displacement vector. For a rigid-body-like rotation of the platelet in shear flow, the displacement vector can be dominated by the large translation/rotation component, so a 2% normalized error in the displacement field does not directly quantify membrane deformation fidelity (strain). The paper's own Fig. 2 and Section 2 indicate the platelet undergoes a Jeffery's orbit, i.e., a large overall rotation. If most of the displacement norm is due to rigid-body rotation, then a 2% error in the displacement could correspond to a much larger relative error in the local deformation (strain), which is what matters for predicting platelet activation. The abstract's phrase 'mode error distribution of approximately 0.5%' is therefore not the same as 'high-fidelity description of platelet deformation.' Furthermore, the claim of generalization to different initial configurations is explicitly deferred in Section 4.2 ('A separate analysis ... will be the subject of a follow-up study'), yet the abstract states the operator predicts deformation 'based on its initial configuration,' suggesting a capability that is not tested. The architecture selection in Section 4.1 is performed on a test set, which can inflate reported accuracy.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a DeepONet surrogate model that maps the initial configuration of a dissipative-particle-dynamics (DPD) platelet and a scalar shear stress value to the platelet's deformed configuration after one Jeffery's orbit period. The training data come from 101 LAMMPS simulations with shear stresses between 50 and 250 Pa. The authors report that the surrogate achieves a maximum relative difference below 2% and a mode of the error distribution of about 0.5% against the LAMMPS ground truth, with a selected architecture of only 6,467 trainable parameters. They also study the influence of network size, the number of training shear-stress cases, and the ability to extrapolate slightly beyond the training range. The stated motivation is to enable a scalable coupling between sub-platelet mechanics and macroscale models of thrombosis.","tokens_in":10237,"tokens_out":7221,"duration_ms":67083,"significance":"If rigorously supported, the paper would demonstrate that a compact neural operator can accurately reproduce the output of a particle-based platelet membrane model, which is a useful component for future multi-scale thrombosis simulations. The paper has several strengths: it reports a genuine out-of-sample extrapolation test, it provides a practical sensitivity study of training-set sparsity, and it identifies a very small architecture (6,467 parameters) that achieves the reported accuracy. However, the central accuracy claim is currently weakened by the use of the test set for architecture selection and by an error metric that does not isolate membrane deformation from rigid-body rotation. With these issues addressed, the contribution would be a modest but solid application of neural operators to a computational biology problem; without them, the quantitative claims are not yet reliable as stated.","major_comments":[{"comment":"The error metric D = ||ξ_pred − ξ_true|| / ||ξ_true|| measures the relative error of the total per-particle displacement vector. Because the platelet undergoes a large rigid-body rotation (a Jeffery's orbit, visible in Fig. 2), a substantial part of ||ξ_true|| is due to rotation and translation, not to membrane deformation. A 2% error in total displacement therefore does not directly quantify the fidelity of the deformation (strain), which is the quantity most relevant to platelet activation. To support the title's 'high-fidelity description of platelet deformation,' the authors should additionally report an error metric that removes rigid-body motion, such as computing the error after subtracting a best-fit rigid rotation and translation, or comparing strain invariants or local bond-length changes.","section":"Section 3.2, Eq. (7)"},{"comment":"The architecture b54 t554 d32 was selected by comparing the relative difference distribution on a test dataset (Fig. 8), and the same test dataset appears to be used in Section 3.2 to report the final errors (Figs. 5 and 6). This creates selection bias: the reported 0.5% mode and 2% maximum are the post-selection performance on the very data used to choose the architecture, so they overstate the expected generalization error. The authors should use a three-way split (training, validation for architecture selection, and a final test set never used in any decision), and report the final accuracy on that independent test set.","section":"Sections 3.2 and 4.1"},{"comment":"The training-set-size study computes the error on the excluded shear-stress cases, but for N = 101 there are no excluded cases, so the N = 101 data point cannot be evaluated on held-out inputs in the same manner. This makes the comparison across N not consistent and weakens the claim that N = 25 is sufficient. The authors should define a fixed set of held-out shear values (e.g., the 10% extremes, or a random 10% of the 101 cases) that is excluded for every N, and report results on that common test set for all N.","section":"Section 4.2"},{"comment":"The model predicts the deformed configuration at a single time point (one Jeffery's orbit period), not the time-resolved trajectory of the membrane particles. The phrase 'captures the dynamics of the membrane particles' (abstract) overstates the temporal scope of the surrogate. Please rephrase to 'predicts the deformed configuration after one orbit period' or explicitly state that time-resolved dynamics are not modeled.","section":"Abstract and Section 3.2"}],"minor_comments":[{"comment":"The unit for 'Blood particle mass' is listed as 's', which is presumably a typo for 'kg'; please correct this in Table 1.","section":"Table 1"},{"comment":"The use of the distribution mode as the headline error statistic is unusual; reporting the median or mean of the relative difference (and its standard deviation) would be more informative and less sensitive to binning choices.","section":"Section 3.2, Fig. 6"},{"comment":"The paper states that TensorFlow was used but does not report the learning rate, optimizer, batch size, or the number of epochs used for the final architecture. Providing these training details (or a link to an open-source repository) would improve reproducibility.","section":"Section 3.1"},{"comment":"The extrapolation study omits only 10% of examples at each end of the training range for the first scenario, so the extrapolation interval is modest. Please state explicitly the actual shear-stress range of the training set and the test points in Pa, rather than only percentages.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of physics.comp-ph and addresses a relevant application of neural operators. The central idea is promising, but the quantitative claims need to be made rigorous with a proper train/validation/test split and a deformation-focused error metric. I believe these issues are fixable within the manuscript's scope; if the authors address them convincingly, the paper could be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a competent proof-of-concept: a small DeepONet (6,467 parameters) reproduces the final configuration of a DPD-simulated platelet under shear with a maximum relative error below 2% on interpolation. That part is believable, and the parity plot and error histograms support it. The sensitivity study is genuinely useful: N=25 training cases are enough, and extrapolation beyond 10% of the training range degrades badly. As a new application of neural operators to platelet deformation, it is a reasonable first step.\n\nThe soft spots are mostly about framing. The error metric in Eq. (7) normalizes by the norm of the total displacement vector, which is dominated by the rigid-body rotation of the platelet (the Jeffery orbit). A 2% error in that vector does not directly bound the error in local deformation or strain, which is what matters for platelet activation. The phrase \"high-fidelity description of platelet deformation\" overstates what the metric shows. Second, the architecture is selected on the test set (Sec. 4.1), so the reported 2% is optimistic; this is common but worth stating. Third, all 101 simulations use the same initial geometry, so the trunk net has never seen a different initial configuration. The abstract says the operator predicts deformation \"based on its initial configuration,\" which suggests a generalization that is explicitly deferred in Sec. 4.2. Fourth, no code or data is released, which limits reproducibility.\n\nNone of this is fatal. The internal logic is sound: the training data come from a published DPD model, the extrapolation test is a real out-of-sample check, and the authors are honest about the follow-up work. The paper would be stronger with a strain-based error metric, one test with a perturbed initial geometry, and a public dataset.\n\nMy recommendation: send it to peer review as a proof-of-concept. It deserves referee time, but the authors should be pushed to temper the \"high-fidelity\" claim and to report error in deformation rather than only displacement.","headline":"Solid proof-of-concept for a DeepONet surrogate of DPD platelet deformation, but the accuracy claim is measured against total displacement, not strain, and generalization to new initial geometries is untested.","tokens_in":10838,"tokens_out":2830,"would_cite":false,"duration_ms":26777,"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":"A neural operator with 6,467 trainable parameters reproduces platelet deformation under shear flow with a maximum relative error below 2% and a mode around 0.5%.","keywords":["neural operator","DeepONet","platelet deformation","dissipative particle dynamics","shear flow","surrogate modeling","thrombosis","multi-scale modeling"],"falsifier":"Take the trained network and evaluate it on a platelet initial configuration not present in the training set, for example a non-ellipsoidal or pre-stressed shape; if the maximum relative displacement error exceeds the reported 2%, the claim that the operator learns the configuration-to-deformation map fails.","tokens_in":9748,"feed_emoji":"🩸","tokens_out":6728,"duration_ms":57993,"temperature":0.7,"pith_summary":"This paper tries to show that a small neural operator can stand in for costly particle-scale simulations of a single platelet deforming in shear flow. The specific claim is that DeepONet, trained on 101 dissipative particle dynamics runs at shear stresses from 50 to 250 Pa, predicts the deformed membrane configuration after one Jeffery's orbit with a mode error around 0.5% and a maximum relative difference below 2%. If the claim holds, sub-platelet deformation can be obtained at nearly zero inference cost and embedded in larger blood-flow models, which would make multiscale thrombosis simulations more tractable.","feed_headline":"A 6,467-parameter network reproduces platelet deformation to 2%","feed_subtitle":"Training on just 25 of 101 shear-flow simulations suffices, opening cheap multiscale thrombosis modeling.","key_machinery":"The machinery is a DeepONet: a branch network encodes the scalar shear stress, a trunk network encodes the initial coordinates of the roughly 18,000 membrane particles, and their inner product followed by a dense layer outputs the deformed coordinates. Training data come from dissipative particle dynamics, in which a 4 by 4 by 1 micron ellipsoidal membrane of harmonically bonded particles is immersed in a Couette flow with parameters inherited from earlier validated blood-platelet models. The network is trained with a mean-squared-error loss on displacement, and the sensitivity analysis uses the resulting error distribution to pick a minimal architecture, to show that 25 shear values suffice, and to probe extrapolation.","core_discovery":"The central claim is that the mapping from initial platelet configuration and shear stress to the deformed platelet configuration after one Jeffery period is learnable by a DeepONet with only 6,467 trainable parameters. Trained on particle dynamics simulation data, the network's predicted particle displacements agree with the simulation to within 2% relative error everywhere, with most particles near 0.5%. The same accuracy is reached with only 25 of the 101 training shear values, and when the extreme 10% of shear values are withheld, the maximum error stays below 10%, while withholding 20% to 30% pushes errors above 20%.","pith_inferences":["Extending beyond the paper: the 0.5% mode error is measured on displacement vectors and is likely dominated by the platelet's overall tumbling motion, so local membrane quantities such as curvature or strain rate could be less accurate than the headline error suggests.","Extending beyond the paper: because all 101 training runs share the same initial ellipsoid, the branch-trunk separation is only tested on the shear-stress input; a natural next test is holding the shear fixed and varying the initial shape, which the paper itself flags as follow-up work.","Extending beyond the paper: the input is a single scalar shear stress, so the trained operator outputs only the end-of-orbit state; feeding a time-resolved shear history could turn the surrogate into a full trajectory predictor, enabling the study of activation kinetics rather than just final deformation.","Extending beyond the paper: if the particle dynamics model's parameters were recalibrated against experimental platelet stiffness data, the same 25-simulation training budget could produce a clinically targeted surrogate for patient-specific shear histories from 4D flow MRI."],"forward_implications":["Within the 50 to 250 Pa shear range, the DeepONet surrogate can replace the particle dynamics simulation for single-platelet deformation at effectively zero inference cost, with maximum relative error below 2%.","Only 25 uniformly spaced shear-stress training cases are needed; generating the other 76 simulations is unnecessary for this accuracy.","The surrogate tolerates modest extrapolation: excluding the extreme 10% of shear values keeps the maximum error below 10%, so small uncertainties in measured clinical shear stress are absorbed.","Because the network size can be kept small without accuracy loss, retraining for new parameter regimes or geometries is computationally cheap.","Embedding this surrogate in a CFD loop would give macroscopic blood-flow models sub-platelet deformation states without running particle simulations at every time step, a route toward multiscale thrombosis modeling."],"supporting_citations":[{"why":"Supplies the DeepONet architecture and the operator universal approximation theorem the surrogate builds on.","marker":"[32]"},{"why":"Defines the dissipative particle dynamics equations and the fluctuation-dissipation weight-function relations used to model blood flow.","marker":"[35]"},{"why":"Provides the validated DPD force-field parameters and blood-platelet interaction model inherited by the simulations.","marker":"[26]"},{"why":"Defines the Jeffery orbit period that sets the simulation and prediction horizon.","marker":"[34]"},{"why":"Provides the particle dynamics simulation software that generated all ground-truth deformation data.","marker":"[33]"},{"why":"Establishes the fluctuation-dissipation condition that fixes the balance between random and dissipative forces in the DPD model.","marker":"[37]"},{"why":"Supplies experimental measurements of platelet passive mechanics used to set the membrane bond elastic constant.","marker":"[38]"},{"why":"Defines the 50 to 250 Pa shear-stress range as the high-shear activation regime the training covers.","marker":"[36]"}],"fun_headline_variants":["Tiny DeepONet maps shear flow to platelet shape within 2%","6,467-parameter network matches platelet deformation within 2%","25 simulations train a network to 2% platelet deformation","Platelet deformation mapped by a 6,467-parameter neural operator"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim stands or falls on whether the dissipative particle dynamics simulation, with its parameter values taken from earlier modeling studies, is an accurate description of a real deforming platelet, and on whether results for the single starting ellipsoid used in all 101 runs transfer to other initial configurations.","fun_headline_variants_meta":{"raw":{"variants":["Tiny DeepONet maps shear flow to platelet shape within 2%","6,467-parameter network matches platelet deformation within 2%","25 simulations train a network to 2% platelet deformation","Platelet deformation mapped by a 6,467-parameter neural operator"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000963,"raw_usage":{"total_tokens":4012,"prompt_tokens":769,"completion_tokens":3243,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":385,"completion_tokens_details":{"reasoning_tokens":3167}},"tokens_in":385,"tokens_out":3243,"duration_ms":20441,"temperature":1.0,"reasoning_tokens":3167,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:03:00.763912+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained network and evaluate it on a platelet initial configuration not present in the training set, for example a non-ellipsoidal or pre-stressed shape; if the maximum relative displacement error exceeds the reported 2%, the claim that the operator learns the configuration-to-deformation map fails.","supporting_citations":[{"cited_title":"Learning nonlinear operators via deeponet based on the uni- versal approximation theorem of operators","cited_arxiv_id":null,"evidence_quote":"Supplies the DeepONet architecture and the operator universal approximation theorem the surrogate builds on."},{"cited_title":"Dissipative particle dynamics: Bridging the gap between atomistic and mesoscopic simulation","cited_arxiv_id":null,"evidence_quote":"Defines the dissipative particle dynamics equations and the fluctuation-dissipation weight-function relations used to model blood flow."},{"cited_title":"Multiscale particle-based modeling of flowing platelets in blood plasma using dissipative particle dynamics and coarse grained molec- ular dynamics","cited_arxiv_id":null,"evidence_quote":"Provides the validated DPD force-field parameters and blood-platelet interaction model inherited by the simulations."},{"cited_title":"The motion of ellipsoidal particles in a viscous fluid","cited_arxiv_id":null,"evidence_quote":"Defines the Jeffery orbit period that sets the simulation and prediction horizon."},{"cited_title":"Statistical mechanics of dissipative par- ticle dynamics","cited_arxiv_id":null,"evidence_quote":"Establishes the fluctuation-dissipation condition that fixes the balance between random and dissipative forces in the DPD model."},{"cited_title":"Quantification of the passive mechanical properties of the resting platelet","cited_arxiv_id":null,"evidence_quote":"Supplies experimental measurements of platelet passive mechanics used to set the membrane bond elastic constant."},{"cited_title":"Results with mechanical cardiac valvular prostheses","cited_arxiv_id":null,"evidence_quote":"Defines the 50 to 250 Pa shear-stress range as the high-shear activation regime the training covers."}],"review_version":1}