{"id":"17084766-5d4c-4b1b-a5c4-eb48cb5a8a91","arxiv_id":"2509.05344","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A gradient-boosting classifier trained on 2,231 polymer-solvent pairs achieves ROC-AUC 0.85 (polymer-level) and 0.91 (solvent-cluster-level) for chemical resistance, with feature analysis highlighting crystallinity, density, and polarity.","lead":"Researchers built a machine learning model that predicts whether a plastic will resist a given organic solvent, using over 2,200 polymer-solvent test results from manufacturer charts. The model performs well in cross-validation and points to polymer crystallinity, density, and solvent polarity as the main drivers, but the study's validation design has caveats.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Labels derived from heterogeneous manufacturer ratings may encode test severity or missingness, not true chemical resistance; the reported AUCs and feature findings are conditional on this target.","rationale":"The reader's weakest assumption identifies target label construction, and I agree that is the load-bearing point. The paper's other weaknesses—label-derived solvent clusters, self-referential feature importance, and missing baselines—would matter less if the target itself were trustworthy; if the target is trustworthy, the polymer-level LOGOCV AUC of 0.85 is at least evidence of learnable signal. Conversely, if the target is biased, no cross-validation scheme fixes it. The proposed test is feasible: the source charts are cited and presumably accessible, and releasing the merged dataset plus source-level ratings would allow a direct audit. I do not think this requires rejecting the paper outright—the modeling pipeline and dataset are potentially useful—but the central claim should be accepted only conditionally on passing this audit. Hence the reader's CONDITIONAL verdict is appropriate; my stress-test pass does not change it.","tokens_in":15750,"tokens_out":5685,"duration_ms":66165,"concrete_test":"Release the dataset with source-level ratings. First, compute inter-source agreement (Cohen's kappa or prevalence-adjusted bias-adjusted kappa) on all polymer–solvent pairs rated by more than one manufacturer. Second, retrain and evaluate the model on (a) only pairs with unanimous 'resistant' or unanimous 'non-resistant' ratings across all sources and (b) a manual audit of 100 randomly sampled pairs re-coded by two independent chemists using a prespecified rubric. If the AUC drops materially (e.g., by more than 0.05) on the high-confidence unanimous subset, or if model errors concentrate in pairs with disagreeing sources or missing ratings, the reported performance is an artifact of the label-construction rule.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the model predicts chemical resistance with ROC-AUC 0.85/0.91 and that crystallinity, density, and solvent polarity are key governing factors. The most load-bearing assumption is the binary label itself (Methods, Datasets). Labels are constructed by (i) labeling only the most stringent resistance rating within a source as 'resistant' (0), treating all milder ratings as 'non-resistant' (1), and (ii) requiring unanimity across sources for a 0, with any single harsh or negative source making the pair 1. This procedure conflates 'resistant under the strictest test and all sources' with 'chemically resistant'. Pairs that were tested only under mild conditions, or that one source rated as conditionally resistant, are coded as non-resistant; pairs with missing ratings in some sources are effectively penalized by the unanimity rule. If the manufacturer charts differ in immersion time, evaluation criteria (weight change, appearance, mechanical properties), or solvent purity, the target is systematically biased by source severity and data completeness. Because the model is trained and evaluated on this target, the reported AUCs and the crystallinity/density/polarity findings may reflect the labeling rule rather than physical chemistry. The manuscript provides no inter-source reliability analysis, no manual audit, and no external experimental validation. This is a correctness risk in the central argument, not a disagreement with consensus.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a machine-learning framework to predict the chemical resistance of polymer–solvent pairs. The authors construct a dataset of 2,231 combinations from six manufacturer chemical-resistance charts, binarize the ratings using the most stringent criterion and unanimity across sources, and train a gradient-boosting decision tree on force-field descriptors, MD-derived polymer properties, a crystallinity label, and a COSMO-RS Flory–Huggins χ parameter. They report ROC-AUC values of 0.85 under leave-one-polymer-out cross-validation and 0.91 under leave-one-solvent-cluster-out cross-validation. The trained model is then applied to 40,971 virtual polymers and 9,828 solvents, and feature analyses are used to claim that polymer crystallinity, polymer density, and solvent polarity are key factors governing chemical resistance. The paper also discusses the χ parameter in relation to a two-step dissolution model for semicrystalline polymers.","tokens_in":16113,"tokens_out":6500,"duration_ms":76255,"significance":"If the dataset and labels are trustworthy, the polymer-level LOGOCV AUC of 0.85 would be a useful contribution, and the physically motivated descriptors (force-field kernel mean embeddings, MD properties, COSMO-RS χ) are a positive feature of the work. The paper also makes concrete, falsifiable predictions for a large virtual chemical space, which is valuable for screening. However, the target-label construction is a major source of uncertainty, the solvent-cluster validation scheme is biased by using outcome-derived clusters, and the feature-importance analysis is largely self-referential. The central claims are therefore not yet established at the level claimed in the abstract.","major_comments":[{"comment":"The binary label definition is load-bearing: combinations are labeled 0 only if they meet the most stringent resistance criterion within a source, and, when multiple sources rate a pair, only if all sources label it resistant. Manufacturer charts differ in immersion time, evaluation metric (weight change, appearance, mechanical properties), and solvent handling, so this rule conflates 'resistant under the strictest available test and all sources' with 'chemically resistant.' Pairs tested only under mild conditions or rated conditionally by one source are coded as non-resistant. No inter-source reliability analysis, manual audit, or external experimental validation is provided. This affects every reported AUC and every feature-interpretation result. Please provide sensitivity analyses (e.g., alternative rating thresholds, excluding ambiguous ratings, per-source agreement statistics) and s","section":"Methods, Datasets"},{"comment":"The solvent clusters used for leave-one-cluster-out CV were constructed by sorting the 132 solvents 'based on the proportion of non-resistant entries' and clustering them into ten groups. This uses the target labels to define the validation groups. Under LOGOCV, the held-out cluster is adjacent in label-proportion space to clusters in the training set, so the model can interpolate along the outcome gradient, and the group structure itself encodes outcome information. The reported 0.91 AUC is therefore not a reliable estimate of generalization to new solvent groups. Please re-run the evaluation with clusters defined from solvent features alone, or with random/stratified folds, and report both estimates. The polymer-level LOGOCV (0.85) is not affected by this specific issue.","section":"Methods, Datasets; Results, Model accuracy (Fig. 3c)"},{"comment":"Crystallinity label and MD density are input features of the resistance model. Figure 5a reports the MIC between the model's predicted non-resistance probability and these same input features, and Figure 6 plots the predicted probability against density. A model given a feature will generally have its output correlate with that feature, so these plots demonstrate that the model uses the features, not that crystallinity or density 'govern chemical resistance.' The abstract's physical claim is not supported by this self-referential analysis. Moreover, the 40,971 virtual polymer resistance predictions are never checked against experiments. Please reframe this as model interrogation, provide an ablation (e.g., retraining without crystallinity or density), or validate on independent experimental data.","section":"Results, Polymer features (Fig. 5a, 5b, 6)"},{"comment":"Figures 7 and 8 analyze the model's predicted non-resistance probability for 9,828 solvents against PE. All of these are model outputs, not measurements, and the polarity/rigidity trends are correlations between the model's output and its input descriptors. Without any experimental check or even a baseline comparison to Hansen solubility parameters, the claim that solvent polarity is a key governing factor of chemical resistance is conditional on the model and on the label construction. The 'like dissolves like' observation is consistent with the model being able to read polarity from the force-field descriptors, but it does not independently confirm a physical mechanism.","section":"Results, Solvent features (Figs. 7–8)"}],"minor_comments":[{"comment":"Typographical and terminology issues: 'F orce field' has a stray space; 'FFKM' is used without definition; 'Kangle' appears both as 'k_angle' and 'Kangle'. Please standardize.","section":"Descriptors"},{"comment":"Figure 3 caption and in-figure labels are somewhat redundant ('2 class classification LOGOCV AUC score for polymer'). Clarify panel labels to avoid confusion.","section":"Results, Model accuracy"},{"comment":"In Figure 7b, the text says nearly all 9,828 solvents are predicted non-resistant, while Figure 8a reports n=558 resistant and n=8,392 non-resistant. Please reconcile these numbers or clarify the threshold used for 'nearly all'.","section":"Results, Solvent features"},{"comment":"The crystallinity classifier is trained on only 27 polymers and applied to 40,971 virtual polymers. The paper acknowledges this in one sentence, but the subsequent interpretation treats predicted crystallinity as reliable. A more prominent caveat and, if possible, a test on a small independent set of polymers would strengthen the analysis.","section":"Results, Polymer crystallinity"},{"comment":"No baseline comparison is made to conventional solubility parameters (e.g., HSP distance or a χ-only classifier). Since the introduction motivates the work as going 'beyond conventional solubility models,' at least one simple baseline would help the reader judge the added value of the proposed descriptors.","section":"Introduction/Results"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's main strengths are the compiled dataset and the physically motivated descriptor set. The target-label construction and outcome-derived solvent clusters are the two most serious risks; both are addressable with sensitivity analyses and revised validation. I do not think the paper should be rejected solely because its conclusions align with textbook solubility theory, but the current presentation overstates the evidence for the feature-interpretation claims. A revision that adds robustness checks and reframes the feature analyses as model interrogation rather than direct physical discovery would be within scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on arXiv:2509.05344. The paper does something genuinely new: it curates 2,231 polymer–solvent chemical resistance labels from manufacturer charts and trains a gradient boosting model on molecular descriptors, achieving 0.85 polymer-level LOOCV AUC. That's a useful baseline for a task (chemical resistance, as opposed to solubility) that has been under-studied. The extension to 40k virtual polymers and 10k solvents is also a reasonable screening exercise, and the semicrystalline chi-parameter discussion aligns with the two-step dissolution model.\n\nThe soft spots are real. The 0.91 solvent-cluster AUC is inflated because the 132 solvents are clustered by the proportion of non-resistant labels before doing leave-one-cluster-out CV. That leaks target information into cluster structure, so the model is essentially tested on within-cluster label variation, not on new clusters. The polymer-level 0.85 is more defensible, but there is no baseline comparison (e.g., HSP-based or simple logistic regression) and no external validation.\n\nThe feature importance analysis is largely circular: they compute MIC between the model's predicted non-resistance probability and input features like crystallinity and density. Of course the model's outputs correlate with its own inputs; this doesn't demonstrate that crystallinity or density 'govern' chemical resistance. To claim that, they would need to vary those features and observe resistance, or at least show predictive power on held-out data.\n\nThe biggest concern is the label itself. They define resistant as meeting the most stringent criterion within any source and unanimous across sources. That means the target is 'passes the strictest test any source applied,' which can encode test severity or data completeness rather than intrinsic chemical resistance. The paper acknowledges the sources differ but provides no inter-source reliability analysis, no manual audit, and no external experimental check. Without that, we can't tell how much signal is physical chemistry and how much is which chart was included.\n\nThis is primarily a screening-tool paper for practitioners in polymer engineering and sustainable materials. It deserves a serious referee because the dataset and the task are valuable, but it needs major revision: fix the cluster leakage by clustering only on descriptors, add baselines and at least a small experimental or literature validation, release the dataset and code, and soften the causal claims. If the authors do that, the contribution would be solid.","headline":"Useful new dataset and a plausible GBDT pipeline for polymer-solvent resistance screening, but the solvent-cluster CV is leaked and the 'key factor' claims are largely self-referential.","tokens_in":16566,"tokens_out":6966,"would_cite":true,"duration_ms":67196,"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 0.91-AUC model predicts polymer resistance to solvents","keywords":["chemical resistance","polymer–solvent compatibility","gradient boosting","force field descriptors","molecular dynamics","Flory–Huggins parameter","crystallinity","solvent polarity"],"falsifier":"Compile a fresh test set of, say, 200 polymer–solvent pairs absent from the 2,231, measure weight change after standardized two-day immersion at fixed temperature, and compare the model's ROC-AUC against those physical measurements. If the model's discrimination collapses—or if its errors line up with which manufacturer's chart a pair came from—the claimed accuracy was an artifact of label construction rather than chemical resistance.","tokens_in":15705,"feed_emoji":"🧪","tokens_out":5236,"duration_ms":53301,"temperature":0.7,"pith_summary":"The paper claims that chemical resistance of a polymer–solvent pair can be predicted from a small set of physics-based descriptors without measuring the pair directly. Training a gradient boosting classifier on 2,231 labels drawn from industrial chemical-resistance charts, the model reaches ROC-AUC 0.85 when entire polymers are left out and 0.91 when solvent clusters are left out. The authors argue that polymer crystallinity and density, together with solvent polarity, are the dominant governing factors, and that the Flory–Huggins interaction parameter is informative mainly for semicrystalline polymers, where solvent penetration begins in amorphous regions. If correct, the result gives a screening tool for replacing persistent fluoropolymers and for designing solvent-resistant materials from computed properties alone.","feed_headline":"A 0.91-AUC model predicts polymer resistance to solvents","feed_subtitle":"Crystallinity, density, and solvent polarity carry the signal, not solubility parameters alone.","key_machinery":"The load-bearing object is the joint feature vector: a kernel-mean-embedded histogram of GAFF2 force-field parameters (atomic mass, charge, van der Waals well depth, bond and angle force constants, and similar terms) for both polymer and solvent, augmented with molecular-dynamics-derived polymer properties (density, radius of gyration, heat capacities, dielectric constants), a binary crystallinity label, and the COSMO-RS Flory–Huggins χ parameter. A gradient boosting decision tree maps this vector to P(non-resistant). The claim is that this combination carries the chemistry that solubility-parameter distances miss.","core_discovery":"On its own terms, the paper establishes a predictive classification of chemical resistance: a polymer–solvent pair is 'resistant' only if every manufacturer chart labels it under the most stringent criterion, and 'non-resistant' otherwise. Using force-field descriptors of both molecules, equilibrium molecular-dynamics properties of the polymer, a crystallinity label, and a COSMO-RS-based Flory–Huggins χ parameter, a gradient boosting decision tree separates the two classes with ROC-AUC 0.85 by polymer and 0.91 by solvent cluster. The authors further show that the learned model encodes physically expected trends: crystalline and denser polymers are predicted more resistant, polar solvents are","pith_inferences":["The label-construction rule (resistant only if all sources agree) likely makes 'non-resistant' include pairs that were merely untested or tested under harsher conditions; the model's probabilities may therefore be better read as resistance confidence than as physical dissolution propensity. This is my inference, not the paper's.","Because the χ parameter failed for amorphous polymers, a natural extension is to add kinetic or swelling-relaxation descriptors (diffusion coefficients, chain-relaxation timescales) for amorphous systems; the paper's two-step picture suggests equilibrium thermodynamics is insufficient there.","The same feature stack could be repurposed to predict continuous outcomes—weight change, swelling ratio, or time-to-failure—rather than binary labels, giving a graded resistance scale for engineering design.","The crystallinity classifier was trained on only 27 polymers yet applied to 40,971 virtual polymers; the downstream density–resistance heatmaps inherit that small-training uncertainty and should be treated as indicative trends, not absolute predictions."],"forward_implications":["Chemical resistance can be screened computationally before synthesis: new polymers can be scored against thousands of solvents from molecular-dynamics descriptors alone.","The identified factors give design rules: raising crystallinity and density, or introducing polar and rigid repeat units, should improve resistance; solvent selection should favor polar, rigid molecules.","χ-based screening is only reliable for semicrystalline polymers; for amorphous or highly crystalline polymers, predictions should rely on other descriptors.","The model provides a benchmark for replacing persistent fluoropolymers with degradable alternatives that match resistance profiles.","A compact descriptor set makes the model's decisions explainable in terms of physical quantities, linking data-driven predictions to established theory."],"supporting_citations":[{"why":"Supplies the manufacturer chemical-resistance charts that define the 2,231 training labels.","marker":"[18–23]"},{"why":"Provides the automated all-atom molecular-dynamics property database used for polymer descriptors.","marker":"[24]"},{"why":"Supplies the 9,828-solvent set used for solvent-side screening and analysis.","marker":"[25]"},{"why":"Defines the kernel mean embedding method that turns force-field parameters into fixed-length descriptors.","marker":"[26]"},{"why":"Prior multitask model that predicts Flory–Huggins χ and solubility, the framework the resistance model builds on.","marker":"[14]"},{"why":"Implements the gradient boosting classifier used as the predictive model.","marker":"[32]"},{"why":"Provides the two-step dissolution theory used to interpret χ behavior in semicrystalline polymers.","marker":"[56]"},{"why":"Supports the swelling-precedes-dissolution mechanism for amorphous and semicrystalline polymers.","marker":"[57]"},{"why":"Shows crystalline domains reduce solvent diffusion, supporting density and crystallinity as resistance factors.","marker":"[41]"},{"why":"Connects solvent polarity to polymer behavior, supporting the polarity finding.","marker":"[47]"}],"fun_headline_variants":["ML predicts polymer solvent resistance with 0.91 AUC","Polymer resistance to solvents decoded by machine learning","Crystallinity, density, polarity guide polymer resistance ML","0.91 AUC model outdoes solubility parameters for polymers","Machine learning maps polymer solvent resistance drivers"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The binary resistance labels are trustworthy: they are assembled from manufacturer charts with different test criteria, and a pair counts as resistant only if every source ranks it under the strictest standard, so 'non-resistant' may partly encode untested pairs or harsher testing rather than true chemical incompatibility.","fun_headline_variants_meta":{"raw":{"variants":["ML predicts polymer solvent resistance with 0.91 AUC","Polymer resistance to solvents decoded by machine learning","Crystallinity, density, polarity guide polymer resistance ML","0.91 AUC model outdoes solubility parameters for polymers","Machine learning maps polymer solvent resistance drivers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000144,"raw_usage":{"total_tokens":950,"prompt_tokens":621,"completion_tokens":329,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":365,"completion_tokens_details":{"reasoning_tokens":254}},"tokens_in":365,"tokens_out":329,"duration_ms":3975,"temperature":1.0,"reasoning_tokens":254,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:58:19.978240+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compile a fresh test set of, say, 200 polymer–solvent pairs absent from the 2,231, measure weight change after standardized two-day immersion at fixed temperature, and compare the model's ROC-AUC against those physical measurements. If the model's discrimination collapses—or if its errors line up with which manufacturer's chart a pair came from—the claimed accuracy was an artifact of label construction rather than chemical resistance.","supporting_citations":[{"cited_title":"Hayashi, J","cited_arxiv_id":null,"evidence_quote":"Provides the automated all-atom molecular-dynamics property database used for polymer descriptors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the 9,828-solvent set used for solvent-side screening and analysis."},{"cited_title":"Kusaba, Y","cited_arxiv_id":null,"evidence_quote":"Defines the kernel mean embedding method that turns force-field parameters into fixed-length descriptors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior multitask model that predicts Flory–Huggins χ and solubility, the framework the resistance model builds on."},{"cited_title":"Pedregosa, F","cited_arxiv_id":null,"evidence_quote":"Implements the gradient boosting classifier used as the predictive model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the two-step dissolution theory used to interpret χ behavior in semicrystalline polymers."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the swelling-precedes-dissolution mechanism for amorphous and semicrystalline polymers."},{"cited_title":"L¨ utzow, A","cited_arxiv_id":null,"evidence_quote":"Shows crystalline domains reduce solvent diffusion, supporting density and crystallinity as resistance factors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Connects solvent polarity to polymer behavior, supporting the polarity finding."}],"review_version":1}