{"id":"98bd92c4-484f-4bc3-ba68-e6a910579bd4","arxiv_id":"2501.02707","paper_version":4,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Bayesian optimization of Martini3 bonded parameters reduces density and radius-of-gyration errors for PE, PMMA, and PS melts at four degrees of polymerization, though the reported errors are in-sample fits rather than out-of-sample predictions.","lead":"This paper uses Bayesian optimization to tune the bonded parameters of coarse-grained Martini3 models for three polymers, targeting all-atom simulation density and radius of gyration. It reports lower errors than default Martini3, but only on the same systems used for fitting, not on unseen chain lengths.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Transferability to any degree of polymerization is asserted but never tested; all 12 reported <10% errors are in-sample fits at the chain length where parameters were optimized.","rationale":"I read the paper in good faith: the BO/CGMD workflow is coherent, the low-dimensional parametrization is sensible, and the per-system error reductions in Figure 5 are plausible. However, the headline claim is not what the data show. The phrase 'validated a CG potential applicable to any degree of polymerization' appears in the abstract and results, and the Introduction frames transferability as the key advantage over prior re-parametrization efforts. The only evidence offered is that each of the 12 systems, after its own optimization, reaches <10% error. That is an in-sample statement. No simulation uses parameters optimized at a different n, no held-out n is evaluated, and no comparison to a uniform-baseline parameter set is made—the latter is explicitly acknowledged as missing. Thus the load-bearing assumption—that the start/middle/end parameters capture transferable physics across chain lengths—is entirely untested. This is not a disagreement with consensus; it is an internal gap between the claim and the experimental design. The 50-PS experimental-density note reinforces the concern by showing the optimized model can be worse than default against experiment, implying the AA target may not be a reliable ground truth and that per-system fitting may chase noise. For these reasons I support the reader's REJECT verdict. The concrete transfer test above would settle the matter: if transfer errors stay under ~10%, the claim could be reinstated; if not, the paper should be reframed as a per-system optimization framework, not a generalizable potential.","tokens_in":15031,"tokens_out":3302,"duration_ms":31407,"concrete_test":"Run a direct transfer test: take the optimized bonded parameter set θ*(n=10) for each polymer (PE, PMMA, PS), build CG topologies for n=20 and n=50 using the same θ* (start/middle/end segments as defined), run the same CGMD protocol, and compute absolute percentage errors in density and Rg against the corresponding AA references. If any transfer error exceeds ~10%, the 'any degree of polymerization' claim is falsified. Also run a uniform single-parameter-set baseline (same bond/angle params for all segments) at n=10 and n=50 to quantify whether the start/middle/end decomposition actually aids transfer.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—'a CG potential applicable to any degree of polymerization' (Abstract and Results)—requires that bonded parameters optimized at one chain length transfer to other n without retraining. This is never tested. All 12 systems (3 polymers × 4 n values) were optimized and evaluated at the same n; Figure 5 reports per-system errors after per-system optimization. The workflow minimizes Eq. (2) separately for each n, so the reported <10% errors are training errors, not predictions for unseen n. The text explicitly acknowledges the missing baseline: 'A direct comparison to a model using a single, uniform parameter set for all segments was not performed but would be valuable future work' (Results, low-dimensional parametrization section). That admission applies equally to cross-n transfer: no θ* optimized at n=3 is ever simulated at n=50. The low-dimensional start/middle/end parameterization is presented as the mechanism enabling transfer, but transferability is assumed, not demonstrated. Without a hold-out test across n, the 'unified model for a given polymer type' is unsupported; the work reduces to per-system fitting. The 50-PS case, where default Martini3 matches experimental density better than the optimized model, further suggests the optimization may overfit to the AA reference rather than improve physical fidelity. The absence of reported final parameter values compounds the issue, preventing independent verification of transfer.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a Bayesian-optimization workflow to refine Martini3 bonded parameters for coarse-grained polymer models. The key methodological idea is a low-dimensional parameterization in which bond length, bond constant, angle magnitude, and angle constant are optimized separately for the start, middle, and end segments of a polymer chain (plus an aromatic bond length for PS). The objective, Eq. (2), is the squared difference between CG and all-atom (AA) density and radius of gyration. For PE, PMMA, and PS at degrees of polymerization n=3, 10, 20, and 50, the authors report absolute percentage errors below about 10% relative to AAMD, show convergence curves and Pareto fronts, and claim that the resulting potential is applicable to any degree of polymerization. However, all reported results are in-sample: each of the 12 systems is optimized and evaluated at the same chain length, and no optimized parameter set from one n is transferred to another n. The central transferability claim is therefore not supported by the present experiments.","tokens_in":15336,"tokens_out":7128,"duration_ms":74202,"significance":"If the transferability claim were established, the workflow would be a practical and comparatively inexpensive way to specialize general-purpose Martini3 topologies for polymer families; the reported convergence in roughly 100 CGMD evaluations per system is plausible and is a useful practical data point. The paper also has clear strengths: the objective is explicitly defined, the optimization pipeline is described in enough detail to be reproduced, five seeds are used for error bars, and the comparison against raw Martini3 is systematic across 12 systems. Those features make the manuscript a potentially useful methodology demonstration. Nevertheless, the paper's principal claimed contribution, a validated 'CG potential applicable to any degree of polymerization,' is not supported by the current experimental design. The reported errors are training errors, and the proposed transfer mechanism is asserted rather than demonstrated. The significance of the work as submitted is therefore limited to per-system fitting, which is a considerably weaker claim than the one made in the Abstract and Results.","major_comments":[{"comment":"The reported errors below ~10% are in-sample training errors, not predictive accuracies. Eq. (2) defines the objective as the squared difference between CG and AA density and radius of gyration, and the workflow minimizes this objective separately for each of the 12 polymer systems. Figure 5 then reports the value of that same objective after optimization. There is no held-out chain length, no held-out parameter set, and no independent validation ensemble. Consequently, the statement that the framework 'reports exceptionally low errors under ~10% for both density and radius of gyration for all the 12 polymers studied' does not support a validated transferable potential; it supports only per-system fitting.","section":"Eq. (2); Results, Pareto Optimal Property Discrepancy Frontier; Figure 5"},{"comment":"The central claim of transferability across degrees of polymerization is not tested. The text states that the start/middle/end parameters 'are intended to be applicable when constructing models for other degrees of polymerization,' and the Abstract and Results assert a validated potential for 'any degree of polymerization.' However, every reported result is for chains where the parameters were optimized at that same n; no parameter set optimized at n=3 is ever simulated at n=50, and no interpolation or extrapolation in n is performed. A required addition is an explicit transfer test: optimize at one n (or on a training subset of n values) and report density and Rg errors at the remaining n values, with the raw Martini3 baseline included for comparison. The paper's own admission that a comparison to a uniform parameter set was not performed is also relevant: such a control is needed to establish that the three-segment parameterization, rather than simply the act of fitting, is responsible for any transferable improvement.","section":"Low-dimensional Parametrization; Results; Abstract"},{"comment":"The optimized parameter values themselves are never reported. Since the paper claims to have developed a CG potential, a table of the final optimized values for all 12 systems (bond lengths, bond constants, angle magnitudes, angle constants, and the aromatic bond length c for PS) is necessary for independent verification, for use by other researchers, and for any meaningful claim of a reusable potential. Without these values, the reported error metrics cannot be checked, the parameters cannot be transferred by readers, and the 'potential' is not actually specified.","section":"Table 1; Results; no table of optimized parameters"},{"comment":"The 50-PS result, where the raw Martini3 model matches experimental density better than the optimized model, exposes a limitation of using AAMD as the sole ground truth. Because the optimization target is the AA density, the workflow can minimize error relative to an AA reference that is itself in tension with experimental data. The authors acknowledge this point, but the central claim of 'accuracy comparable to AAMD' is then less informative unless experimental comparisons are also reported and the choice between AA-calibrated and experimentally calibrated targets is discussed. At minimum, the paper should state explicitly that the optimized potential is a fit to the AA model, not necessarily to experiment.","section":"Results, Pareto Optimal Property Discrepancy Frontier; 50-PS discussion"}],"minor_comments":[{"comment":"The sentence 'Furthermore, this low-dimensional parametrization approach balances computational efficiency with the need to capture key topological features of the CG polymer' appears twice, at the end of the first and second paragraphs of the section; one copy should be removed.","section":"Low-dimensional Parametrization"},{"comment":"The log marginal likelihood is missing a leading minus sign in the first term; as written, maximizing the expression would not correspond to the standard Gaussian process marginal likelihood, and the sign error would affect hyperparameter optimization.","section":"Eq. (5)"},{"comment":"The PS start/end bond constant upper bound is listed as 90000, which is likely a typographical error (perhaps 9000), and the row for PE/PMMA angle constants uses k_b where k_phi is clearly intended.","section":"Table 1"},{"comment":"The notation for z_k^(m) and l_k^(m) is introduced too briefly; the superscripts, subscripts, and the min over candidate points should be defined precisely, and the text should state explicitly that Eq. (12) is an expanded Monte Carlo estimator of the same quantity defined in Eq. (11).","section":"Eqs. (11)-(12)"},{"comment":"The error bars represent standard deviation over five seeds, but the text does not specify what varies between seeds (initial Latin Hypercube design, BO stochasticity, or MD initial velocities); this should be clarified so that the uncertainty is interpretable.","section":"Figure 5"},{"comment":"The manuscript does not state whether code, input files, example topologies, or final optimized parameter sets will be made available; a data and code availability statement is needed for reproducibility.","section":"Methods and reproducibility"}],"recommendation":"major_revision","confidential_remarks":"I recommend major revision rather than rejection because the methodological core is sound and the missing transfer validation and parameter reporting could, in principle, be addressed in a revision. However, the current text should not be accepted as-is: the Abstract and Results make a strong transferability claim that the experiments do not support. If the authors cannot provide a cross-n transfer test or explicitly remove the 'any degree of polymerization' claim, the paper should not be accepted. The 50-PS experimental-density caveat should also be given more weight in the conclusions, as it bears on the choice of optimization target."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know upfront. First, the low-dimensional start/middle/end parametrization for polymer bonded terms is a genuinely clever, practical idea. Second, the paper's central claim—that the resulting CG potential works for any degree of polymerization—is not tested anywhere in the manuscript. All twelve reported <10% errors come from models optimized and evaluated at the same chain length, so they are in-sample fits to the objective function.\n\nThe BO workflow itself is standard but competently executed: they use an ARD-SE GP surrogate, q-EHVI acquisition, and a reasonable 20-run Latin hypercube initial design. The idea of reducing a chain's bonded parameters to three segment types is useful and could help others. The paper also honestly notes the missing uniform-parameter baseline and the 50-PS experimental density issue.\n\nThe load-bearing gap is the transferability claim. The objective in Eq. (2) is the squared difference between CG and AA density and radius of gyration, so after optimization the errors are by construction the training values. There is no simulation at n=50 using parameters optimized at n=3, or vice versa. Without that, the 'any degree of polymerization' statement is an assertion. The lack of final parameter values, code, or data makes independent verification impossible. The 50-PS result, where default Martini3 matches experiment better than the optimized model, suggests the AA reference may not always be the right target.\n\nThis is a useful methods paper in waiting, but the current evidence only supports 'BO can fit these three polymers at four chain lengths.' I would not cite the transferability claim, and I'd want to see a hold-out test across chain lengths before taking it seriously. Still, the core idea is sound enough that I'd send it to peer review with a strong request for that experiment and for releasing the parameter sets. It's a desk-reject only if you think the authors can't or won't address the gap.","headline":"Clever start/middle/end parametrization for Martini3 bonded terms, but the 'any degree of polymerization' claim is asserted, not tested—all reported errors are in-sample fits.","tokens_in":15901,"tokens_out":2250,"would_cite":false,"duration_ms":23429,"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":"This paper claims that refining only the bonded parameters of Martini3 topologies by Bayesian optimization—twelve to fifteen numbers per polymer—reproduces all-atom density and radius of gyration within about 10 percent error for all…","keywords":["coarse-grained molecular dynamics","Martini3 force field","Bayesian optimization","multi-objective optimization","polymer topology","density","radius of gyration","force field refinement"],"falsifier":"Take the optimized start/middle/end parameters for, say, 10-PE and use them unchanged in a CGMD simulation of 50-PE; compare the computed density and radius of gyration to the corresponding all-atom values. If the absolute percentage errors exceed roughly 10 percent, the claim that the potential applies to any degree of polymerization is falsified.","tokens_in":14850,"feed_emoji":"⚛️","tokens_out":6930,"duration_ms":61466,"temperature":0.7,"pith_summary":"Coarse-grained simulations trade atomistic detail for speed, and the general-purpose Martini3 force field often misses the molecular-scale structure of a specific polymer. This paper argues that a small, targeted retuning of Martini3's bonded parameters—bond lengths, bond constants, angle magnitudes, angle constants, plus one aromatic bond length—can recover all-atom accuracy. Using Bayesian optimization with a multi-output Gaussian process and the q-EHVI acquisition function, the authors tune only the start, middle, and end segments of a polymer chain, cutting the parameter space from 139 to 15 dimensions for a 20-unit polystyrene chain. They report that for all 12 polymer systems (PE, PMMA, and PS at degrees of polymerization 3, 10, 20, and 50), the optimized coarse-grained simulations match all-atom density and radius of gyration within roughly 10 percent error, and they claim the resulting CG potential transfers to any degree of polymerization. If true, this gives a cheap way to build accurate coarse-grained models for domain-specific polymer applications.","feed_headline":"Bayesian optimization cuts coarse-grained polymer error to under 10%","feed_subtitle":"A 12-to-15 parameter retune of Martini3 reproduces all-atom density and radius of gyration for PE, PMMA, and PS.","key_machinery":"The load-bearing mechanism is a low-dimensional parametrization of the CG polymer topology: only the bonded parameters of the first, middle, and last segments (bond lengths $b_0$, bond constants $k_b$, angle magnitudes $\\Phi$, angle constants $k_\\Phi$, and, for aromatic polystyrene, an aromatic bond length $c$) are optimized, giving 12 or 15 design variables. The optimizer is a multi-output Gaussian process with an automatic-relevance-determination squared-exponential kernel and a linear model of coregionalization that jointly predicts density and radius of gyration, driven by the q-EHVI acquisition function, which balances exploration and exploitation while tracking the Pareto front between the two objectives. This machinery maps expensive CGMD runs to a tractable black-box optimization problem and is what lets the paper claim accuracy comparable to all-atom MD at CGMD cost.","core_discovery":"The central claim is that a Martini3 topology refined by multi-objective Bayesian optimization can serve as a coarse-grained potential that is both accurate and transferable across chain lengths. The authors define molecular topology as the bonded parameters within a fixed CG mapping, and they compress the full bonded parameter set of a polymer chain into three groups—start, middle, and end segments—so that optimizing a 20-unit polystyrene chain requires only 15 numbers instead of 139. All-atom NPT simulations provide the reference values for density and radius of gyration; each CG parameter set is evaluated by running Martini3 simulations, and a Gaussian-process surrogate built with a linear model of coregionalization predicts both objectives so that q-EHVI can select the next batch of candidates. Iterating roughly 50 rounds (100 CGMD runs) drives the absolute percentage errors below about 10 percent for all 12 systems, with five independent seeds per polymer. The paper therefore proposes that the optimized start/middle/end parameters constitute a CG potential applicable to any degree of polymerization, a claim the authors put forward as a benchmark for future CG model building.","pith_inferences":["Editorial inference: The transferability claim is not yet tested because every reported result optimizes and evaluates at the same chain length; the assertion that the start/middle/end parameters work for any degree of polymerization is an extrapolation from that data.","Editorial inference: Because PMMA's density error grows with chain length and the paper attributes this to Martini3's non-bonded polar/nonpolar bead description, optimizing bonded terms alone may hit a ceiling for polar polymers; a combined bonded-plus-nonbonded refinement would be a testable extension.","Editorial inference: The start/middle/end compression implicitly assumes that polymer ends differ from the interior in a way that is independent of chain length; that assumption could be checked by computing full per-monomer parameters at several degrees of polymerization and comparing them to the optimized three-segment values."],"forward_implications":["If the optimized potentials transfer across chain lengths, one short optimization run per polymer class replaces per-length reparameterization.","The same loop can be pointed at experimental observables instead of all-atom reference values, since the objective function is built from abstract target properties.","The protocol's roughly 50-iteration convergence (about 100 CGMD runs, 3–11 hours on a single GPU) makes domain-specific Martini3 refinement affordable for routine materials screening.","Adding extra parameters to the start/middle/end scheme extends the approach to copolymers, branched chains, or polymers with side-group chemistry beyond PE, PMMA, and PS."],"supporting_citations":[{"why":"Defines the original Martini CG force field and its bead-mapping and interaction matrix, the baseline this work refines.","marker":"13"},{"why":"Martini3 force field and its small-molecule parameterization supply the default CG topologies whose bonded terms are optimized.","marker":"21–23"},{"why":"GROMACS is the MD engine used for both all-atom reference simulations and coarse-grained evaluation runs.","marker":"34"},{"why":"Maximum-projection Latin hypercube design seeds the initial 20 CGMD runs used to train the surrogate.","marker":"55"},{"why":"q-EHVI is the acquisition function that selects candidate parameter sets during the multi-objective optimization.","marker":"56"},{"why":"Gaussian-process regression is the surrogate model that predicts density and radius of gyration from the low-dimensional parameters.","marker":"76"},{"why":"Linear model of coregionalization builds the multi-output covariance that encodes correlations between the two objectives.","marker":"79,80"},{"why":"GAFF defines the all-atom bonded and non-bonded interactions used to generate reference AAMD data.","marker":"64,65"},{"why":"Martinize2 and Vermouth convert all-atom structures into Martini3 CG topologies for the three polymers.","marker":"67"},{"why":"J-OCTA prepares the all-atom structure files for the 12 polymer chains used as simulation inputs.","marker":"63"}],"fun_headline_variants":["Bayesian tuning makes coarse-grained polymers near all-atom accurate","Machine-learning retune of Martini3 slashes CG polymer error to 10%","Smart parameter tweaks let coarse-grained simulations match all-atom","Optimized Martini3 potential: polymer accuracy with CG speed","Bayesian refinement of Martini3 cuts polymer simulation error tenfold"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's transferability claim rests on the untested premise that parameters optimized for start, middle, and end segments at one degree of polymerization work at other chain lengths without retraining, since all validation is done at the same chain length used for optimization.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian tuning makes coarse-grained polymers near all-atom accurate","Machine-learning retune of Martini3 slashes CG polymer error to 10%","Smart parameter tweaks let coarse-grained simulations match all-atom","Optimized Martini3 potential: polymer accuracy with CG speed","Bayesian refinement of Martini3 cuts polymer simulation error tenfold"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00015,"raw_usage":{"total_tokens":1244,"prompt_tokens":1039,"completion_tokens":205,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":655,"completion_tokens_details":{"reasoning_tokens":113}},"tokens_in":655,"tokens_out":205,"duration_ms":2785,"temperature":1.0,"reasoning_tokens":113,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:06:05.956461+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the optimized start/middle/end parameters for, say, 10-PE and use them unchanged in a CGMD simulation of 50-PE; compare the computed density and radius of gyration to the corresponding all-atom values. If the absolute percentage errors exceed roughly 10 percent, the claim that the potential applies to any degree of polymerization is falsified.","supporting_citations":[{"cited_title":"J., Risselada, H","cited_arxiv_id":null,"evidence_quote":"Defines the original Martini CG force field and its bead-mapping and interaction matrix, the baseline this work refines."},{"cited_title":"R., Gul, E","cited_arxiv_id":null,"evidence_quote":"Maximum-projection Latin hypercube design seeds the initial 20 CGMD runs used to train the surrogate."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Martinize2 and Vermouth convert all-atom structures into Martini3 CG topologies for the three polymers."},{"cited_title":"(Springer Singapore, Singapore, 2016)","cited_arxiv_id":null,"evidence_quote":"J-OCTA prepares the all-atom structure files for the 12 polymer chains used as simulation inputs."}],"review_version":1}