{"id":"ec43492e-6072-4af6-bb6e-ad210563eb7f","arxiv_id":"2411.14555","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A DeepONet surrogate reproduces finite element post-burn contraction fields across multiple wound shapes with R2 = 0.99 and CPU/GPU speedups of 128x and 235x.","lead":"This paper trains a DeepONet neural operator to reproduce finite element simulations of post-burn skin contraction, reporting R2 = 0.99 on unseen wound shapes made from convex combinations of training shapes. A generalist should read it because it shows how neural surrogates could turn slow burn-wound simulations into near-instant predictions for treatment planning.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Shape encoding is not injective on the convex test set: Section 3.3 admits the quadruple does not uniquely determine test shapes, so the trunk cannot distinguish some distinct inputs; reported R2 cannot establish generalization unless FE outputs are shown to be insensitive to the missing shape…","rationale":"The paper is a useful, well-documented surrogate-modelling study with released code and data, and the architecture comparison is informative. My concern targets the single condition on which the advertised generalization rests: the trunk's shape representation must be sufficient for the FE displacement field. The authors themselves state in Section 3.3 that it is not unique for the convex test set. Because the model output is a function of that quadruple (plus coordinates and parameters), any two test shapes that collide on the quadruple but differ in FE output make the test set ill-posed; the aggregate R2 cannot detect this. Nothing in the paper quantifies the sensitivity of the morphoelastic model to shape variation at fixed quadruple, nor how the quadruple is extracted from Eq. (9). The proposed computational check would settle the issue: construct or find quadruple-colliding convex combinations and compare their FE fields. If they differ materially, the claim of generalization to convex combinations is unsupported; if they are nearly identical, the encoding is in fact sufficient and the reader's condition is met. Since this is exactly the reader's weakest assumption, my stress test does not change the conditional verdict, but it sharpens the acceptance criterion.","tokens_in":19487,"tokens_out":9306,"duration_ms":96369,"concrete_test":"Take two weight triples (alpha1,alpha2,alpha3) and (alpha1',alpha2',alpha3') in Eq. (9) that produce different convex-combination boundaries but identical (ycut,xm,ym,xcut) values; generate FE displacement fields for both and compare them, especially on the wound boundary and at the time of maximal contraction. Also scan the 150 existing convex-test simulations for quadruple collisions and report the maximum FE discrepancy among collision groups. If non-negligible discrepancies exist, the supervised problem is ill-posed and the R2 should be recomputed per collision group; if the FE outputs are nearly identical, or if no such pairs occur in the test set, the encoding suffices and the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the trained DeepONet generalizes to convex combinations of the three training shapes. The mechanism by which it could do so is the four-number shape descriptor (ycut, xm, ym, xcut) fed to the trunk network (Section 2.2, Fig. 3). For the convex test set, however, Section 3.3 states explicitly that this quadruple \"no longer uniquely define[s] the initial wound shapes.\" Two different convex combinations can therefore be mapped to the same trunk input while having different finite-element displacement fields, and the model is structurally unable to distinguish them. The paper gives no procedure for computing the quadruple from a convex combination, and no evidence that the unrepresented part of the shape has negligible effect on the morphoelastic solution. The headline metrics R2=0.99 and aRRMSE=0.075 are computed by pooling 18,035,821 points over 150 simulations; any small number of quadruple collisions, or collisions whose differences occur mainly in low-displacement interior regions, would be masked by this aggregation. The test therefore cannot, as designed, support the generalization claim; good aggregate metrics are also compatible with interpolation within the training-shape family or with dominance of near-zero displacements enforced by the sine augmentation. This is an internal limitation admitted by the authors, not a disagreement with an external consensus.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper trains a DeepONet as a surrogate for a two-dimensional morphoelastic finite-element model of post-burn dermal contraction. The branch network takes five patient-specific biological parameters, while the trunk network takes space-time coordinates together with a four-number wound-shape descriptor (ycut, xm, ym, xcut); sine augmentation is used to enforce homogeneous boundary conditions. The model is trained on 750 FE simulations of rectangle, rhombus, and ellipse wound shapes and tested on 150 FE simulations whose geometries are convex combinations of the three training shapes. On a pooled test set of about 18 million points the authors report R2 = 0.99437, aRRMSE = 0.07497, and aRelErr = 0.03429, with CPU/GPU speedups of about 128x and 235x, respectively. The one-year extension requires retraining on new FE simulations with tend = 365 days; the paper reports this honestly and compares two retraining data-sampling strategies.","tokens_in":19840,"tokens_out":5327,"duration_ms":51871,"significance":"If the generalization claim holds, this is a useful step toward geometry-flexible neural surrogates for wound-healing simulations, with potential clinical planning applications. The paper has clear strengths: the code and data are publicly available, the architecture ablation in Table 2 is informative, the one-year extrapolation failure is reported rather than hidden, and the speedup measurements are concrete. However, the main claim of generalization to convex combinations rests on a shape encoding that the paper itself admits is not injective for the test set, so the headline metrics do not yet establish the stated generalization. The work is therefore promising but requires additional evidence or an architectural change.","major_comments":[{"comment":"The trunk input quadruple is claimed in Section 2.2 to uniquely determine the initial wound shape, but Section 3.3 explicitly states that for the convex test set the quadruple \"no longer uniquely define[s] the initial wound shapes.\" Since the trunk is the only channel carrying shape information, two different convex combinations with the same quadruple are mapped to the same input and thus to the same predicted displacement field. The paper does not describe how the quadruple is computed for a convex combination, and it provides no evidence that the unrepresented part of the shape has negligible effect on the FE displacement field. Consequently, the pooled R2 = 0.994 in Table 2 does not by itself establish generalization to unseen geometries; it is also compatible with interpolation within the training-shape family or with dominance of small-displacement points. Please quantify the severity of the non-injectivity: identify distinct convex combinations sharing a quadruple, compare their FE displacement fields, and report per-simulation or per-collision-group metrics. Alternatively, replace the four-number descriptor with an injective shape representation, such as values of a level-set or signed-distance function at fixed sensor points.","section":"Section 3.3 and Section 2.2"},{"comment":"The reported R2, aRRMSE, and aRelErr are computed by pooling about 18 million points from only 150 FE simulations, in which the points are highly correlated within each simulation across space and time. No uncertainty intervals, per-simulation score distributions, or effective sample sizes are given. A small number of shape collisions, or collisions whose differences occur mainly in low-displacement interior regions, would be masked by this aggregation. Please report the distribution of scores over the 150 simulations (e.g., per-run R2), and ideally identify how many distinct quadruples occur and how many collisions exist.","section":"Section 4.1 and Eqs. (10)-(12)"},{"comment":"The abstract states that the model \"provided reliable predictions over an extended period of up to one year,\" but Section 4.3 shows that the model trained on t in [0,100] fails after day 100 unless it is retrained on new FE simulations with tend = 365. The abstract and conclusions should state clearly that the one-year predictions require additional FE data and fine-tuning, and that the original model does not extrapolate temporally. This is an honest limitation already present in the body, but the abstract currently overstates the temporal scope.","section":"Abstract and Section 4.3"}],"minor_comments":[{"comment":"The sentence \"The latter uniquely determines the initial wound shape at t = 0\" should be restricted to the three training shapes, since Section 3.3 later revokes this property for the convex test set.","section":"Section 2.2"},{"comment":"The convex combination in Eq. (9) requires the parametrizations r_p(s) to share a common parameter interval and orientation, but the paper does not specify these parametrizations. Please add the explicit form of r_p(s) for the rectangle, rhombus, and ellipse used in the experiments.","section":"Section 3.3"},{"comment":"The relative error metric uses displacements rounded to one decimal place; please clarify the rationale and state whether the reported values are sensitive to this rounding.","section":"Section 4.1, Eq. (12)"},{"comment":"The text labels the retraining strategies as \"scenario 1\" and \"scenario 2,\" but the figure legend describes them by the number of FE simulations and time steps; please align the terminology so the reader can map the text to the figure.","section":"Section 4.3, Fig. 10"},{"comment":"There is a minor typo in Section 3.3: \"The convex test set it created\" should read \"The convex test set is created.\"","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The main load-bearing concern is the non-injective shape descriptor, which is internal to the paper's own claims and not a disagreement with an external consensus. I believe the issue is addressable within the scope of a revision, for example by adding a collision analysis or by changing the shape encoding, so I recommend major revision rather than rejection. The paper's code/data availability and honest reporting of the one-year retraining requirement are positive signals."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nHere's the short version: this is a competent, honest surrogate-model paper, and it should go to peer review, but the headline generalization claim is shakier than the R2 and the abstract suggest. The stress-test note is right, and the authors themselves flag the weak point.\n\nWhat's new and good: it's the first DeepONet for post-burn contraction, moving past the fixed-geometry MLPs in [6,9]. Adding the shape quadruple to the trunk and sine augmentation are sensible, standard modifications, and the ablation table (Tab. 2) shows they matter. The code and data are on GitHub/4TU, which is exactly what this subfield needs. The speedup comparison is careful, including the important caveat that their speedups are lower than the earlier MLPs because they predict the full displacement field.\n\nThe soft spots, in order of seriousness.\n\nOne: the shape encoding is not injective on the convex test set. Section 3.3 literally says the quadruple 'no longer uniquely define[s] the initial wound shapes.' Two distinct convex combinations can map to the same trunk input, so the model is structurally blind to some differences between test shapes. The paper does not quantify how often this happens or whether the FE displacement is sensitive to the missing geometry. The pooled metrics over 18 million points could easily mask a small number of collisions. This does not sink the paper, but it does mean the strong generalization claim is not established by the evidence as presented.\n\nTwo: the architecture comparison in Table 2 appears to select the best model using the same convex test set. There is no mention of a separate test set held out from that selection, so the reported R2=0.99 is optimistic. This is fixable with a nested split or a fresh test set.\n\nThree: there's an internal inconsistency. Section 4.1 says 'most of the spatial coordinates in the training set are located at the wound boundaries,' but Section 3.1 says 20 spatial coordinates are uniformly sampled per time step. Minor, but should be corrected.\n\nFour: the one-year claim in the abstract needs qualification. The paper honestly shows the 100-day model fails to extrapolate, then retrains on new FE data with tend=365. That's a fine investigative result, but the abstract's wording overstates it.\n\nThere are other minor things (no uncertainty quantification, parameter ranges narrow), but they don't change the picture.\n\nWho this is for: researchers working on neural surrogates for soft-tissue mechanics, especially wound healing. It is a legitimate incremental application paper with reproducible artifacts. My recommendation: send it to peer review, but with a firm request to address the shape-encoding limitation and to rerun the evaluation on a test set that was not used to choose the architecture. If the authors can show that non-injectivity rarely matters, the central claim will hold; if not, the claim should be softened.","headline":"A solid, honest surrogate-model paper with reproducible code, but the generalization claim rests on a shape encoding the authors admit is not injective; worth peer review with a request to fix the evidence.","tokens_in":20361,"tokens_out":3173,"would_cite":true,"duration_ms":30399,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65M60","68T07","92C50"],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural operator predicts post-burn contraction on unseen wound shapes with R^2 = 0.99 and 128–235x speedup.","keywords":["operator learning","DeepONet","post-burn contraction","surrogate model","finite element method","morphoelasticity","wound healing","neural network surrogate"],"falsifier":"Take two convex combinations with the same quadruple $(y_{\\mathrm{cut}}, x_m, y_m, x_{\\mathrm{cut}})$ but different weights, run the finite-element solver on both, and compare the two displacement fields; if the DeepONet returns nearly identical predictions while the solver's outputs differ by more than the reported error band, the shape encoding is not carrying the deformation-relevant geometry and the generalization claim would be falsified.","tokens_in":19298,"feed_emoji":"🩹","tokens_out":8349,"duration_ms":69227,"temperature":0.7,"pith_summary":"This paper tries to establish that a deep operator network (DeepONet) — a neural network that learns maps between function spaces — can replace slow finite-element simulations of the two-dimensional morphoelastic model that governs post-burn skin contraction. The authors train on three basic wound shapes (rectangle, rhombus, ellipse) and test on convex combinations of those shapes that were never seen during training, reporting $R^2 = 0.99$, an average relative root mean squared error of $0.075$, and speedups of about 128-fold on CPU and 235-fold on GPU. If the claim holds, clinicians could obtain patient-specific predictions of dermal displacement and wound-relative surface area in seconds rather than minutes, which is what interactive treatment planning would need.","feed_headline":"Burn-contraction surrogate hits R² of 0.99 and runs 235x faster","feed_subtitle":"Trained on three wound shapes, it reproduces finite-element displacement on convex blends 128x faster than the solver.","key_machinery":"The load-bearing object is the DeepONet architecture with its branch/trunk split. The branch is a multi-layer perceptron that maps five scalar parameters — cell diffusion constant, chemotactic constant, signaling diffusion constant, cell differentiation rate, and half-maximum cell division enhancement rate — to coefficient vectors $b_1,\\dots,b_{2p}$. The trunk is another multi-layer perceptron that maps $(t,x,y)$ together with the wound-shape quadruple $(y_{\\mathrm{cut}}, x_m, y_m, x_{\\mathrm{cut}})$ to basis values $c_1,\\dots,c_p$; the outputs are $\\hat{u}_1 = \\sum_i b_i c_i$ and $\\hat{u}_2 = \\sum_i b_{i+p} c_i$. Sine augmentation then enforces the boundary conditions exactly by multiplying $\\hat{u}_1$ by $\\sin(\\pi x/x_l)\\cos(\\pi y/(2y_l))$ and $\\hat{u}_2$ by $\\sin(\\pi y/y_l)\\cos(\\pi x/(2x_l))$. The quadruple is the part that lets the trunk adapt its learned basis to a new wound geometry, and it is what carries the claimed generalization.","core_discovery":"The central claim is that the solution operator of the post-burn contraction PDE system can be learned by a DeepONet whose branch network takes five patient-specific biophysical parameters and whose trunk network takes the evaluation point $(t,x,y)$ plus a four-number encoding $(y_{\\mathrm{cut}}, x_m, y_m, x_{\\mathrm{cut}})$ of the initial wound shape; the dot product of the two network outputs gives the displacement field $(u_1,u_2)$. A sine-augmentation layer multiplies the raw outputs by sinusoidal factors, so the homogeneous boundary conditions on the rectangular domain hold exactly rather than approximately. Trained on 750 finite-element simulations run to $t=100$ days, the model reaches $R^2 = 0.99437$, $\\mathrm{aRRMSE} = 0.07497$, and $\\mathrm{aRelErr} = 0.03429$ on an 18-million-point convex test set. Extending the horizon to one year by adding 50 simulations with $t_{\\mathrm{end}}=365$ days and sampling 30 time steps per run brings the predicted relative surface area at $t=365$ to within roughly half a percent of the finite-element target. From the ablation comparisons the paper concludes that initial-shape information must enter through the trunk, and that adding new finite-element runs improves generalization more than adding extra time samples from existing runs.","pith_inferences":["Editorial inference: the convex test set is a low-dimensional slice of wound geometry, so the strongest reading of the claim is tied to shapes built from the same three prototypes; a genuinely harder test would use asymmetric or non-convex wounds, which the paper itself lists as future work.","Editorial inference: because the paper notes that the quadruple no longer uniquely identifies shapes in the convex set, feeding an image of the initial wound (the convolutional route the authors mention) would test whether the network has truly learned geometry or merely a smooth interpolation in the four-parameter encoding.","Editorial inference: the comparison of sampling strategies suggests a data-acquisition rule for clinical practice — include more patients with fewer samples per patient — that could be tested by training the same DeepONet on datasets with matched total sample size but different patient counts."],"forward_implications":["Burn-contraction simulation can be reduced to a learned operator: given five patient parameters and a four-number wound-shape descriptor, the full two-dimensional displacement field over 100 days is obtained in seconds rather than minutes.","Because the trunk receives the shape encoding, the same trained network applies to any convex combination of the three training shapes without retraining, supporting generalization beyond the exact geometries in the training set.","Extending the time horizon to one year requires adding simulations that reach $t=365$; with 50 added runs sampled at 30 time steps, the predicted relative surface area stays within about 0.6 percent of the finite-element reference in the worst case.","Initial wound-shape information must be fed to the trunk network: moving it to the branch degrades $R^2$ from 0.99 to 0.90, and omitting sine augmentation along with shape information gives an aRRMSE around 0.32.","Adding new finite-element runs to the training set improves one-year predictions more than adding more time samples from existing runs, suggesting data diversity across patients matters more than density per patient."],"supporting_citations":[{"why":"Supplies the morphoelastic PDE model whose finite-element solutions are the training targets.","marker":"[24]"},{"why":"Defines the DeepONet architecture and the universal approximation theorem on which the surrogate is built.","marker":"[29]"},{"why":"Previous 2D neural-network surrogate on a fixed wound shape; the baseline for speedup and geometric flexibility comparisons.","marker":"[9]"},{"why":"Earlier Bayesian-trained machine learning model for post-burn contraction; the prior work this study extends.","marker":"[6]"},{"why":"Gives the operator universal approximation result that the branch/trunk decomposition relies on.","marker":"[2]"},{"why":"Early use of neural networks to hard-enforce boundary conditions, cited to justify the sine augmentation trick.","marker":"[26]"}],"fun_headline_variants":["DeepONet predicts burn contraction 235x faster than FEM","AI surrogate for burn scars: 99% accurate, 128-235x speedup","Neural operator models post-burn contraction at R² 0.99","Burn contraction AI: R² 0.99, 235x faster on GPU","DeepONet speeds burn-contraction prediction 235x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the four-number shape encoding $(y_{\\mathrm{cut}}, x_m, y_m, x_{\\mathrm{cut}})$ carries enough geometric information for the trunk network to distinguish the wound shapes that matter, even though the convex test set contains different shapes that this quadruple does not uniquely determine; if the encoding is insufficient, the high scores may reflect interpolation among the three training shapes rather than true generalization.","fun_headline_variants_meta":{"raw":{"variants":["DeepONet predicts burn contraction 235x faster than FEM","AI surrogate for burn scars: 99% accurate, 128-235x speedup","Neural operator models post-burn contraction at R² 0.99","Burn contraction AI: R² 0.99, 235x faster on GPU","DeepONet speeds burn-contraction prediction 235x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000515,"raw_usage":{"total_tokens":2577,"prompt_tokens":1096,"completion_tokens":1481,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":712,"completion_tokens_details":{"reasoning_tokens":1392}},"tokens_in":712,"tokens_out":1481,"duration_ms":9040,"temperature":1.0,"reasoning_tokens":1392,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:08:28.308308+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take two convex combinations with the same quadruple $(y_{\\mathrm{cut}}, x_m, y_m, x_{\\mathrm{cut}})$ but different weights, run the finite-element solver on both, and compare the two displacement fields; if the DeepONet returns nearly identical predictions while the solver's outputs differ by more than the reported error band, the shape encoding is not carrying the deformation-relevant geometry and the generalization claim would be falsified.","supporting_citations":[{"cited_title":"A mathematical model for the simulation of the formation and the subsequent regression of hyper- trophic scar tissue after dermal wounding","cited_arxiv_id":null,"evidence_quote":"Supplies the morphoelastic PDE model whose finite-element solutions are the training targets."},{"cited_title":"DeepONet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators","cited_arxiv_id":null,"evidence_quote":"Defines the DeepONet architecture and the universal approximation theorem on which the surrogate is built."},{"cited_title":"High-speed predictions of post-burn con- traction using a neural network trained on 2d-finite element simulations","cited_arxiv_id":null,"evidence_quote":"Previous 2D neural-network surrogate on a fixed wound shape; the baseline for speedup and geometric flexibility comparisons."},{"cited_title":"A Bayesian finite- element trained machine learning approach for predicting post-burn contraction","cited_arxiv_id":null,"evidence_quote":"Earlier Bayesian-trained machine learning model for post-burn contraction; the prior work this study extends."},{"cited_title":"Chen and H","cited_arxiv_id":null,"evidence_quote":"Gives the operator universal approximation result that the branch/trunk decomposition relies on."},{"cited_title":"Lagaris, A","cited_arxiv_id":null,"evidence_quote":"Early use of neural networks to hard-enforce boundary conditions, cited to justify the sine augmentation trick."}],"review_version":1}