{"id":"10d8f8ed-e1cd-49ef-969b-aed7719840a6","arxiv_id":"2505.11685","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A physics-augmented machine learning model trained on 8,900 numerical contact simulations predicts viscoelastic pull-off force and work-to-pull-off across Tabor parameter, material spectrum, preload, and unloading rate.","lead":"This paper builds machine learning models that quickly predict how strongly a rigid sphere sticks to a soft, stretchy material when it is pulled off, using data from detailed contact simulations. The models combine a physics formula with data and could help engineers design grippers and soft robots that control adhesion in real time.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Broad-band extrapolation (n=0.2) is validated only against XPB, not BEM, so a key part of the 'wide range' accuracy claim lacks ground truth.","rationale":"The reader identified the lack of experimental validation of the BEM ground truth as the weakest assumption. I agree that is a limitation, but the more immediately load-bearing gap is internal: the paper's showcase extrapolation to n=0.2 is compared only with XPB, and the PA-ML architecture uses XPB as a feature. In the absence of BEM or experimental data in that region, the claim of 'accurate predictions in a wide range of conditions' cannot be distinguished from 'the model reproduces its own analytical feature.' The interpolation claims in the sampled regime are well supported: high R2 on BEM holdout, reproducible data/code, and physically expected saturation behavior. This does not invalidate the paper; it narrows the proven scope. Consequently, the conditional verdict stands, but the conditions should explicitly include BEM or experimental checks in the extrapolation region.","tokens_in":20231,"tokens_out":7415,"duration_ms":78615,"concrete_test":"Generate a small set of fresh BEM simulations at the previously inaccessible extrapolation points (e.g., n=0.2, µ=3.24, k=0.1, δl=73, br_u=10^1, 10^3, 10^5) using a higher computational budget, adaptive time stepping, or a second independent BEM/FEM solver, and compare PA-ML predictions to these BEM outputs. If log10(bΓeff) deviations exceed the test-set MSE band, or if the depth-dependent Tabor trend in Figure 9 changes, restrict the claimed validity range. If the BEM truly cannot converge at those points, run at n=0.3 and δl=50 and use a convergence study in n and δl to bound the extrapolation error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is a fast, accurate surrogate across broad-band viscoelastic materials, with n=0.2 cited as silicone-like (Section 3.1). In Section 4.1 and Figure 6, the paper states that for n≤0.2, δl≥73, µ≥3.24 the BEM 'fails to determine the pull-off force' at reasonable computational cost, so no training data exist there. The evidence for generalization in this regime is agreement with the XPB dashed curves, not with BEM outputs. XPB is itself a fitted analytical model from [20], and this paper does not validate XPB for n=0.2 in this regime against an independent solver or experiment. Moreover, because the PA-ML model receives XPB's output as an input feature, agreement with XPB in extrapolation is partly built into the model architecture: it shows the model can mirror XPB, not that XPB is correct. Thus, the broad-band, deep-indentation, JKR-like corner of the parameter space—arguably the most physically relevant for PDMS—is exactly where the claimed accuracy is unsupported. The interpolated regime (BEM holdout tests, Tables 1-4) is fine; the problem is the scope of the claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes physics-augmented machine learning (PA-ML) surrogates for the normalized pull-off force (effective surface energy) and the work-to-pull-off of a rigid sphere unloaded from a viscoelastic adhesive half-space. Training targets are generated with a Boundary Element Method (BEM) that combines a Lennard-Jones traction law, a Boltzmann history integral, and a modified power-law creep compliance (Eqs. 1-4). The model inputs are the Tabor parameter, the power-law exponent, the modulus ratio, the indentation depth, and the unloading rate; the PA-ML variant adds the analytical XPB prediction (Eq. 5) as an additional feature. The authors compare linear regression, regression trees, random forest, and XGBoost using five-fold cross-validation, reporting R-squared values of about 0.9995 for effective surface energy and 0.9956 for work-to-pull-off, and they use the trained PA-ML model to explore rate-, depth-, and Tabor-parameter-dependent adhesion. The paper also provides open data and code on Zenodo and GitHub.","tokens_in":20446,"tokens_out":5382,"duration_ms":60192,"significance":"If the accuracy claim is understood as accuracy relative to the BEM model, this is a useful and reproducible surrogate- modelling contribution: the train/test split on BEM targets makes the interpolation evaluation non-circular, the data and code are openly available, and the predicted JKR and DMT limits in Figure 9 are physically sensible. The value of the surrogate lies in the region where XPB is invalid and BEM is expensive. However, the paper's broader claims extend to a region where BEM has no ground truth and where the only reference is XPB, which is also used as an input feature; that part of the claim is not independently established and should be re-scoped or supplemented.","major_comments":[{"comment":"The broad-band extrapolation claim for n=0.2, delta_l >= 73, and mu >= 3.24 is not validated by BEM, because the authors state that in this regime the BEM fails to determine the pull-off force at reasonable computational cost. The only reference used in Figure 6 is the XPB analytical model. Since Section 3.2 feeds the XPB output as an input feature to the PA-ML model, agreement with XPB in this regime is partly built into the model architecture and does not independently establish accuracy. Please either validate this regime with an independent numerical method or experiment, or explicitly rephrase the claims as interpolation plus consistency with XPB, rather than as validated 'wide range' accuracy.","section":"Section 4.1, Figure 6"},{"comment":"The ground truth throughout the paper is entirely numerical: there are no experimental measurements and no convergence or uncertainty quantification for the BEM solutions. The abstract's statement that the model 'properly predicts' pull-off in soft materials like silicones and elastomers therefore goes beyond what the manuscript demonstrates, because the BEM itself inherits the assumptions of the Lennard-Jones law, the Boltzmann integral, and the modified power-law compliance. Please add an explicit statement that accuracy is established only with respect to the BEM model, or add an experimental benchmark.","section":"Section 2, Appendix A"},{"comment":"The XPB feature is not parameter-free: it depends on the fitted constants alpha ~ pi/9 and the crack-velocity relation v-hat = 2.887 r-hat_u^1.171 taken from reference [20]. Consequently, in the extrapolated regime the PA-ML model inherits any systematic error in XPB. The manuscript should state this limitation explicitly and, ideally, test the sensitivity of the PA-ML predictions to plausible variations of these constants, so that the reader can judge how much of the reported generalization comes from physics guidance rather than from a specific fitted analytical model.","section":"Section 3.2, Eq. (5), Tables 2 and 4"}],"minor_comments":[{"comment":"The text contains small language issues: 'an Hertzian profile' should be 'a Hertzian profile', and 'silicons' should be 'silicones'.","section":"Abstract"},{"comment":"There is a typo in 'inidentation depth' in the conclusions, and the word 'approachesd' appears in the final sentence of Section 5.","section":"Section 5"},{"comment":"The paper states in Section 3.2 that the total number of data samples 'does not exceed 8505', while Section 5 says '8921 samples generated'. This numerical inconsistency should be reconciled.","section":"Section 3.2 vs Section 5"},{"comment":"Equation (A.6) contains an unmatched parenthesis and the notation 'XbGij' is not defined; please define the influence-matrix notation and correct the formula.","section":"Appendix A.1, Eq. (A.6)"},{"comment":"The axis labels in Figure 4 are difficult to read, with superscripts such as '^ru', '^n', and '^delta_l' appearing in a garbled typeset form; please regenerate the figure with clear mathematical notation.","section":"Figure 4"},{"comment":"The comparison of PA-ML versus ML for work-to-pull-off uses the XPB-predicted effective surface energy as an additional feature. Since the physical connection between Gamma_eff and the work-to-pull-off is not discussed, please add a remark on why this feature is expected to help, rather than merely describing the MSE improvement.","section":"Section 4.2, Tables 3 and 4"}],"recommendation":"major_revision","confidential_remarks":"I agree with the reader's main concern: the extrapolation claim in the n=0.2, deep-indentation, JKR-like corner is supported only by agreement with XPB, which is also an input feature. The interpolation results on BEM holdouts are convincing, and the open data and code are valuable. I would encourage the editor to request a revision that either adds an independent numerical or experimental validation in that regime or carefully re-scopes the claims in the abstract and conclusions. The paper is suitable in scope for the journal if revised accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid surrogate-model paper, not a physics breakthrough. It trains tree-based regressors (XGBoost, Random Forest) on roughly 8,900 BEM simulations of a rigid sphere unloading from a viscoelastic adhesive half-space, and adds the authors' own XPB analytical model as an extra input feature. On BEM holdout data the surrogates are excellent: R² ≈ 0.9995 for effective surface energy and 0.9956 for work-to-pull-off, with JKR and DMT limits respected. The data and models are openly available on Zenodo/GitHub, which is a real plus.\n\nWhat's actually new: prior ML work in adhesion focused on flat pillar geometry and elastic behavior; this is the first ML surrogate for viscoelastic Hertzian pull-off across Tabor parameter, spectrum exponent, modulus ratio, preload, and unloading rate. The physics augmentation does reduce MSE and helps in sparse-data regions. That part is genuine.\n\nThe soft spots are about scope, not derivation. The ground truth is entirely numerical—Lennard-Jones traction, Boltzmann history integral, modified power-law compliance. No experiments, no error bars on the BEM data. If that physical model is wrong, the ML accuracy does not transfer to real materials.\n\nThe bigger issue is the claim about the broad-band, deep-indentation, JKR-like corner (n ≈ 0.2, δl ≥ 73, μ ≈ 3.24). The paper says BEM cannot produce pull-off there at reasonable cost, so no training data exist. The evidence for generalization is agreement with XPB dashed curves, not with BEM. But XPB is also an input feature of the PA-ML model, so the model can mirror XPB without independently validating XPB. The abstract's 'wide range of conditions' overstates what is actually supported. The interpolated regime is fine; the extrapolated corner is unvalidated.\n\nMinor: the dataset count is inconsistent (8,505 vs 8,921). Also, the abstract's 'real-time predictions in soft materials like silicons and elastomers' goes beyond what a BEM-trained surrogate can claim without experiments.\n\nBottom line: this deserves a serious referee and probably acceptance after revision. The authors should either add experimental validation for a few parameter combinations or sharply reframe the claims to the BEM regime. The physics augmentation is useful but not an independent check. For someone working on adhesive-contact surrogates, the dataset and trained models are worth citing.","headline":"A solid surrogate-model paper with genuinely useful data and models, but the wide-range generalization claim leans on an extrapolation that is checked only against the authors' own analytical model.","tokens_in":21032,"tokens_out":2484,"would_cite":true,"duration_ms":23562,"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 physics-augmented machine-learning model trained on boundary-element simulations can predict the pull-off force and detachment work of a rigid sphere pulled from a viscoelastic substrate in milliseconds, covering regimes where analytic…","keywords":["viscoelastic adhesion","pull-off force","physics-augmented machine learning","broad-band viscoelasticity","boundary element method","Tabor parameter","work-to-pull-off","Hertzian contact"],"falsifier":"Run colloid-probe retraction experiments on a well-characterized silicone at fixed geometry, with independent viscoelastic characterization, over preloads from shallow to deep and retraction rates spanning the rubbery-to-glassy transition; if the measured normalized pull-off force or work-to-pull-off deviates from the PA-ML predictions by more than the stated test error in the low-Tabor, low-preload regime, the simulation-to-reality transfer claimed by the model is falsified. A cheaper internal check is to recompute a fresh random set of parameter combinations with an independent boundary-element implementation and compare against the surrogate.","tokens_in":19951,"feed_emoji":"🧲","tokens_out":12735,"duration_ms":115012,"temperature":0.7,"pith_summary":"The paper claims that a physics-augmented machine-learning model (PA-ML) can act as a fast, accurate surrogate for boundary-element simulations of a rigid sphere being unloaded from a broad-band viscoelastic adhesive substrate. Trained on roughly 8,921 simulations and given the analytic XPB estimate of effective surface energy as an additional input, the model predicts normalized pull-off force and work-to-pull-off from five dimensionless parameters: Tabor parameter, power-law spectral exponent, modulus ratio, indentation depth, and unloading rate. Compared with purely data-driven ML, adding the analytic guidance cuts mean squared error by about 60 percent for effective surface energy and improves extrapolation to regions where no training data exist. This matters because the trained model runs in under five milliseconds, while similar boundary-element cases can take hours, and it exposes a depth-dependent Tabor effect that analytic theories miss.","feed_headline":"Physics-boosted ML predicts sticky contact pull-off in milliseconds","feed_subtitle":"A model trained on simulations and guided by an analytic contact law covers regimes where both fail.","key_machinery":"The machine is a supervised regression pipeline on tabular data with five dimensionless physical inputs, a boundary-element dataset generated from the intermolecular force-separation law, the Boltzmann history integral, and a modified power-law creep compliance, plus one extra input, the analytic XPB model's effective surface energy. XPB is a closed-form integral that extends steady-state viscoelastic crack theory to broad-band materials; feeding its output to the regressor lets the model learn only the residual between the analytic approximation and the numerical truth. This is what makes the surrogate fast, small, and interpretable, and it also gives the model a physically consistent anchor when extrapolating beyond the training set, for example to power-law exponents for which simulations failed to converge.","core_discovery":"The central finding is that the analytic XPB prediction, used as a physics-guidance feature, lets tree-based regressors correct the analytic error rather than learn the whole contact law from data. This physics augmentation reduces cross-validated MSE from about $1.45 \\times 10^{-4}$ to $5.75 \\times 10^{-5}$ for effective surface energy and improves agreement with XPB in untrained regions, while remaining faithful to boundary-element test data where XPB is invalid, namely at low Tabor parameter, low indentation depth, and high unloading rate. Using the trained model to sweep parameters, the paper identifies a depth-dependent Tabor effect: at shallow indentation depths the effective surface energy grows with the Tabor parameter, while at deep indentations a larger Tabor parameter shifts behavior toward the short-range-adhesion (JKR) limit and lowers adhesion. For work-to-pull-off, the area under the tensile part of the unloading curve, no analytical model exists; the PA-ML model predicts a bell-shaped dependence on unloading rate and an interaction with the power-law exponent such that the exponent raises detachment work at low rates and lowers it at high rates.","pith_inferences":["The paper's accuracy claims are about reproducing simulated boundary-element physics; transferring the surrogate to real silicones or elastomers would require calibration or experimental validation, since no laboratory measurements appear in the training or test data.","The success of feeding the XPB output as a feature suggests a general recipe: any closed-form contact approximation can be used to augment ML surrogates for other outputs, such as contact area, friction, or energy release rate, in regimes where the approximation is partly valid.","The predicted depth-dependent Tabor effect is directly testable with colloid-probe or atomic-force-microscope retraction experiments at fixed rate and varying preload; if confirmed, it implies that shallow-indentation grippers benefit more from high-surface-energy materials than deep-indentation contacts do."],"forward_implications":["Contact simulations that took up to about 9.8 hours per case can be replaced by sub-millisecond predictions, making real-time adhesion control or design-loop optimization feasible for grippers, climbing robots, and soft-material manufacturing.","The model spans the transition between short-range (JKR-like) and long-range (DMT-like) adhesion in viscoelastic contacts, including low-preload, low-Tabor, and high-rate regimes where the analytic XPB model fails, so it can serve as the general estimator there.","Adhesion strength in soft viscoelastic contacts can be tuned through preload and indentation depth as well as retraction rate and material spectrum; the predicted depth-dependent Tabor effect is a new design lever.","The work-to-pull-off predictions provide an energy-based detachment criterion where no analytical formula exists, useful for energy budget calculations in soft robotics and adhesive interfaces."],"supporting_citations":[{"why":"Supplies the analytic XPB model whose output is the physics-augmentation feature and the extrapolation baseline.","marker":"[20]"},{"why":"Provides the steady-state viscoelastic crack-propagation theory that the XPB model extends to broad-band materials.","marker":"[28]"},{"why":"Defines the work-to-pull-off integral and documents indentation-depth effects that motivate the ML surrogate.","marker":"[34]"},{"why":"Supplies the force-separation law used as the interfacial traction in the boundary-element simulations.","marker":"[66]"},{"why":"Provides the modified power-law creep-compliance model that defines the broad-band viscoelastic substrate.","marker":"[68]"},{"why":"Supplies the overlapping-triangles spatial discretization used in the boundary-element solver that generates the training data.","marker":"[72]"},{"why":"Gives the short-range (JKR-like) elastic adhesive contact solution that sets one limiting regime the model must recover.","marker":"[10]"},{"why":"Provides the gradient-boosted tree algorithm used as one of the best-performing regressors in the comparison.","marker":"[74]"}],"fun_headline_variants":["Physics-boosted ML nails sticky contact pull-off force","Analytic-guided ML corrects pull-off force in viscoelastic contacts","Physics augmentation halves error in ML pull-off prediction","Hybrid model predicts pull-off force fast without heavy numerics","ML with physics features speeds up adhesive contact prediction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the boundary-element simulations, built on the force-separation law, the Boltzmann history integral, and the modified power-law compliance, faithfully represent real viscoelastic adhesive pull-off; the paper validates against its own numerical solver and the analytic XPB model, with no experimental data.","fun_headline_variants_meta":{"raw":{"variants":["Physics-boosted ML nails sticky contact pull-off force","Analytic-guided ML corrects pull-off force in viscoelastic contacts","Physics augmentation halves error in ML pull-off prediction","Hybrid model predicts pull-off force fast without heavy numerics","ML with physics features speeds up adhesive contact prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001293,"raw_usage":{"total_tokens":5324,"prompt_tokens":1039,"completion_tokens":4285,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":655,"completion_tokens_details":{"reasoning_tokens":4204}},"tokens_in":655,"tokens_out":4285,"duration_ms":31518,"temperature":1.0,"reasoning_tokens":4204,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:49:47.741636+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run colloid-probe retraction experiments on a well-characterized silicone at fixed geometry, with independent viscoelastic characterization, over preloads from shallow to deep and retraction rates spanning the rubbery-to-glassy transition; if the measured normalized pull-off force or work-to-pull-off deviates from the PA-ML predictions by more than the stated test error in the low-Tabor, low-preload regime, the simulation-to-reality transfer claimed by the model is falsified. A cheaper internal check is to recompute a fresh random set of parameter combinations with an independent boundary-element implementation and compare against the surrogate.","supporting_citations":[{"cited_title":"Maghami, Q","cited_arxiv_id":null,"evidence_quote":"Supplies the analytic XPB model whose output is the physics-augmentation feature and the extrapolation baseline."},{"cited_title":"Persson, E","cited_arxiv_id":null,"evidence_quote":"Provides the steady-state viscoelastic crack-propagation theory that the XPB model extends to broad-band materials."},{"cited_title":"Johnson, J","cited_arxiv_id":null,"evidence_quote":"Supplies the force-separation law used as the interfacial traction in the boundary-element simulations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the modified power-law creep-compliance model that defines the broad-band viscoelastic substrate."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the overlapping-triangles spatial discretization used in the boundary-element solver that generates the training data."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the short-range (JKR-like) elastic adhesive contact solution that sets one limiting regime the model must recover."}],"review_version":1}