{"id":"44dc2d83-3d67-4bb5-8a14-e1806c59b187","arxiv_id":"2502.07400","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An explainable multimodal ML analysis of 40 aqueous-dispersion CNT fibers finds small uniform aggregates and long defect-free nanotubes respectively drive strength and conductivity, while density drives stiffness.","lead":"This paper applies an explainable machine learning pipeline, combining non-negative matrix factorization and SHAP, to identify which multiscale features control the strength, conductivity, and stiffness of carbon nanotube fibers. It is worth reading as a practical example of how interpretable ML can suggest processing targets for nanomaterials when datasets are small.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"NMF rank and RF hyperparameters are selected on the same 40 samples, so the high LOOCV R2 and the SHAP-identified 'crucial' aggregate features may be selection artifacts rather than stable structure-property relationships.","rationale":"The paper's central claim is that the EMML pipeline, specifically NMF-derived basis weights plus SHAP, reveals that small, uniformly distributed aggregates drive fracture strength and long CNT effective lengths drive electrical conductivity. For this claim to hold, the model must have genuinely learned these relationships from the multiscale inputs rather than from noise or sample-specific quirks. The most load-bearing concern is statistical: with 40 samples, NMF rank and random forest hyperparameters are both selected using the same target variable and the same LOOCV procedure, and the reported R2 and SHAP values come from that same selection. This creates a selection-bias loop: the chosen 7-component NMF decomposition and the tuned RF are the ones that best fit the 40 samples, so the subsequent SHAP attributions are optimistically biased and may not reflect stable structure-property relationships. The reader's weakest assumption about NMF rank selection and overfitting captures exactly this issue, so I agree with the reader's assessment. I do not see a different, stronger concern: the data description and methods are clear, the qualitative findings align with prior CNT literature, and the use of SHAP for local/global interpretation is sensible once a valid model is available. However, without a nested or independent validation that the identified basis components and SHAP trends persist out-of-sample or under resampling, the central claims should be treated as hypotheses. The conditional verdict is therefore appropriate, and my analysis does not change it.","tokens_in":10653,"tokens_out":3042,"duration_ms":31491,"concrete_test":"Run a nested LOOCV: for each held-out sample, fit NMF with ranks 1-40 on the other 39 samples, select the rank by inner LOOCV on those 39 samples, tune RF hyperparameters by inner CV, and predict the held-out sample. If the resulting nested R2 is substantially below the values in Figure 3(a) (or near zero), the reported performance and the SHAP-based attributions are dominated by selection bias. As a complementary check, permute the target variables and repeat the same rank-selection/SHAP pipeline on noise; if the rank-7 R2 and the sign/stability of the Basis000/001/004/006 SHAP curves are statistically indistinguishable from the permutation null, the claimed structure-property relationships are not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2 states that 'the number of factors in NMF was optimized based on predictive performance' (Figure 3a), and Figure S2c indicates that random forest hyperparameters were likewise tuned using the same Leave-One-Out cross-validation. With only 40 samples, selecting both the NMF rank and the random-forest hyperparameters against the same target variable on the same samples, then reporting LOOCV R2 and SHAP values from that same model, introduces optimistic selection bias. The reported R2 in Figure 3(a) may therefore overstate the model's predictive value, and the SHAP attributions in Figures 5 and 8 are derived from a model whose feature set was not independently validated. NMF solutions are also non-unique under scaling and component ordering, and no stability analysis or external test set is provided. Consequently, the central claims that 'small, uniformly distributed aggregates are crucial for improving fracture strength' and 'long effective lengths are significant for electrical conductivity' rest on a model whose generalization to new dispersions or fibers has not been demonstrated. This is not an objection to the use of NMF or SHAP per se, but to the absence of any uncertainty quantification or validation step that would separate genuine structure-property relationships from idiosyncratic patterns in the 40 training samples.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an Explainable Multimodal Machine Learning (EMML) pipeline for CNT fibers, combining multimodal characterization data (DCS aggregate size distributions, FIR effective length, Raman IG/ID, morphological features, and processing conditions) with NMF-based feature extraction, random forest regression, and SHAP-based interpretation. Using 40 CNT fiber samples from a prior experimental study, the authors select a 7-component NMF decomposition of DCS distributions based on LOOCV R2, then report SHAP global importance and local interaction analyses. They conclude that fracture strength is improved by small, uniformly distributed aggregates and a small amount of medium-sized aggregates with wide distribution widths; electrical conductivity is enhanced by long effective CNT lengths and a narrow distribution of small aggregates; and Young's modulus is dominated by density and fiber morphology. Thresholds for IG/ID and effective length are also proposed as processing targets.","tokens_in":10829,"tokens_out":3171,"duration_ms":31525,"significance":"If the results are robust, the paper offers a useful template for combining factor analysis with explainable AI in multiscale materials informatics, and it connects specific, measurable dispersion-state features (NMF basis weights) to macroscopic fiber properties. The comparison of PCA and NMF on the DCS distributions is a clear strength, and the local SHAP interaction plots provide more actionable information than standard feature-importance rankings. The claimed structure-property relationships are plausible and consistent with prior CNT fiber literature, and the proposed thresholds could guide dispersion-process optimization. However, the significance is conditional on the stability and independent validation of the 7-component NMF model and the random forest regressions; the current evidence is not yet sufficient to support the strong mechanistic language used in the abstract and Section 3.4.","major_comments":[{"comment":"The number of NMF components and the random forest hyperparameters are both selected by maximizing LOOCV R2 on the same 40 samples, and the reported R2 values and SHAP attributions are then derived from that same selected model. This procedure leaks information from the left-out folds into the model-selection step, so the LOOCV R2 in Figure 3(a) and the corresponding actual-versus-predicted plots likely overstate predictive performance. Because the central claim that the 7-component decomposition yields meaningful aggregate populations depends on this selection, the authors should either perform a nested cross-validation (selecting NMF rank and hyperparameters within each training fold) or provide an external test set, and they should report confidence intervals for the R2 values and for the NMF component weights across resamples.","section":"Section 2; Section 3.2; Figure 3(a); Figure S2(c)"},{"comment":"NMF solutions are not unique: they are subject to scaling and component-ordering ambiguities and can depend on the initialization and on small perturbations of the input data. The paper does not report any stability analysis, such as running NMF with multiple random seeds, bootstrapping the 40 DCS distributions, or clustering the resulting basis vectors across runs. Since the interpretation in Figure 8 assigns distinct physical meanings to Basis000, Basis001, Basis004, and Basis006, the authors need to demonstrate that these components are stable under resampling and re-initialization; otherwise the SHAP-based attribution of strength and conductivity changes to specific basis components could be an artifact of a particular factorization.","section":"Section 3.2; Section 3.3; Figures 4 and 8"},{"comment":"The claim that SHAP interaction values should be interpreted only above a \"threshold of 10%\" is introduced without a formal justification, and the threshold is applied after scaling the left axis to approximately 5% of the maximum target value. This is load-bearing for the threshold statements about IG/ID ~30 and effective length ~2000 nm, because different scaling choices could change which points fall above the threshold. The authors should either derive the threshold from a principled criterion (for example, a noise floor estimated from repeated model fits) or present the sensitivity of the qualitative conclusions to the chosen threshold.","section":"Section 3.3; Figures 7 and 8"},{"comment":"The language in the abstract and conclusion states that small, uniformly distributed aggregates are 'crucial' for improving fracture strength and that long effective lengths are 'significant factors' for conductivity, implying causal or mechanistic status. What the analysis actually establishes is predictive association within one set of 40 samples. The authors themselves note that XAI 'does not directly clarify mechanisms' in Section 2, but the later discussion nonetheless draws causal conclusions such as 'the presence of these aggregates introduces subtle structural heterogeneity, which helps to reduce crack and promote the redistribution of stress.' These mechanistic interpretations should be clearly flagged as hypotheses consistent with the SHAP analysis, not as demonstrated mechanisms, unless supported by targeted experiments or physical simulations.","section":"Section 3.4; Figure 9"}],"minor_comments":[{"comment":"The term 'Negative Matrix Factorization' is used in the Abstract; the correct standard term is 'Non-negative Matrix Factorization' (also used elsewhere in the paper).","section":"Abstract; Introduction"},{"comment":"The text discusses the IG/ID ratio, but the right-axis labels in Figure 7 (and the corresponding supplementary figures S10-S12) show 'ID / IG'. Please make the notation consistent.","section":"Figure 7 caption and Section 3.3"},{"comment":"The text states that the coefficient of determination reached its maximum at 7 components for all three target variables, but the figure shows only the general trend; please indicate explicitly whether the maximum is attained at exactly 7 for each property or whether 7 is a compromise across the three targets.","section":"Section 3.2, Figure 3(a)"},{"comment":"The sentence 'Figure 5 provides a global evaluation of how explanatory variables to regression predictions' appears to be missing a verb (e.g., 'contribute to'); please rephrase.","section":"Section 3.3, Figure 5"},{"comment":"The unit 'mg/cm' is used for linear density, which is acceptable if intended as tex-like linear density, but it should be defined clearly in the main text or captions to avoid confusion with volumetric density.","section":"Supplementary Figures S10-S12"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the journal's scope and presents a promising methodological framework. However, the validation gap regarding model selection on the same 40 samples is substantial for a paper whose main contribution is the interpretation of structure-property relationships. I would support publication after the authors provide nested cross-validation or an external validation set, a stability analysis of the NMF decomposition, and a more cautious framing of the mechanistic conclusions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid, clearly written case study of an interpretable-ML workflow for CNT fibers, but the validation is not strong enough to support the causal and threshold claims in the abstract. The stress-test concern lands: the NMF rank and the random forest hyperparameters are both selected using the same 40 samples and the same LOOCV R2, so the reported R2 and the SHAP attributions carry selection bias. That doesn't make the paper worthless—the NMF-versus-PCA comparison is a genuinely useful demonstration, and the qualitative findings (small uniform aggregates help strength, long effective length helps conductivity) are plausible and consistent with prior work—but it does mean the paper should be read as hypothesis-generating rather than hypothesis-testing.\n\nWhat's new: applying NMF to DCS size distributions to get interpretable basis components, then using SHAP on a random forest to link those components to fiber properties, and doing this for aqueous-dispersion CNT fibers. The NMF decomposition into small/medium/large aggregate components, with the demonstration that PCA gives wavy uninterpretable components, is a nice, reproducible illustration of why NMF can be better for non-negative distribution data. The paper is also honest about the procedure: it states that the number of NMF factors was optimized based on predictive performance, and the supplementary shows hyperparameter tuning via LOOCV.\n\nThe soft spots are real and concentrated in validation. 40 samples is small for a random forest with many features plus SHAP. Selecting both the NMF rank and RF hyperparameters on the same LOOCV R2 overstates predictive performance; there is no nested CV, no bootstrap, no stability check on the NMF solution (which is non-unique under scaling and ordering), and no independent test set. The SHAP findings are read from the same fitted model, so they are partly post hoc descriptions of the training data. The 10% threshold for 'meaningful contribution' is ad hoc. Also, Figure 3(b) is labeled 'test data' but it is actually LOOCV predictions, which is a misleading label. These issues don't invalidate the direction of the findings, but they do mean the specific thresholds (e.g., effective length > 2000 nm) should be treated as tentative.\n\nWho gets value: materials informatics researchers looking for a worked example of NMF+SHAP on real experimental data, and CNT fiber experimentalists who want a quick map of which dispersion features to control. It deserves a serious referee—the method is sensible and the dataset is real—but it needs major revision on validation before publication. The authors should add nested CV, bootstrap or repeated CV with different NMF ranks, and clearly label LOOCV results as such. Ideally they'd release data and code.","headline":"A clearly written NMF+SHAP case study on CNT fibers whose central causal claims are undercut by selection-on-the-same-40-samples validation, but which is still worth a serious referee.","tokens_in":11426,"tokens_out":2161,"would_cite":false,"duration_ms":19447,"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":"Using NMF-extracted dispersion features and SHAP, this paper shows that small, uniformly distributed aggregates improve fracture strength while long effective CNT lengths improve electrical conductivity, with density dominating Young's…","keywords":["carbon nanotubes","CNT fibers","explainable AI","SHAP","nonnegative matrix factorization","dispersion state","structure-property relationships","multimodal machine learning"],"falsifier":"Prepare fibers from dispersions whose DCS aggregate-size distributions are deliberately varied (for example, by controlled sonication or filtration) while holding other variables fixed, and check whether fracture strength and conductivity track the SHAP-identified basis weights; alternatively, bootstrap the 40-sample dataset to see whether the same seven basis components and their SHAP signs reappear.","tokens_in":10387,"feed_emoji":"🧵","tokens_out":5184,"duration_ms":44085,"temperature":0.7,"pith_summary":"The paper proposes an Explainable Multimodal Machine Learning (EMML) pipeline that combines multiscale characterization, nonnegative matrix factorization (NMF) for feature extraction, Random Forest regression, and SHAP analysis to identify which structural features of carbon nanotube (CNT) fibers actually drive macroscopic properties. Applied to 40 aqueous-dispersion CNT fibers, it finds that fracture strength is governed by small, uniformly sized aggregates in the dispersion, electrical conductivity by long effective CNT lengths and a narrow aggregate-size distribution, and Young's modulus by fiber density and morphology. The claim matters because it converts opaque structure-property correlations into concrete processing targets for bottom-up nanomaterial manufacturing.","feed_headline":"Small aggregates set CNT fiber strength, long tubes set conductivity","feed_subtitle":"A multimodal ML pipeline using NMF and SHAP pinpoints dispersion-state targets for stronger, more conductive nanotube fibers.","key_machinery":"NMF (nonnegative matrix factorization) converts each DCS aggregate-size distribution into weights on seven nonnegative basis distributions, each with a clear peak position and width grouping (small, medium, large). These basis weights are the interpretable features that Random Forest regression consumes along with FIR effective length, Raman I_G/I_D, density, linear density, and process variables. SHAP interaction values then decompose each prediction into contributions per feature, yielding both global importance rankings and local threshold/trend relationships. The paper's methodological claim is that for nonnegative, multi-peaked distribution data, NMF yields sparser and more physically interpretable components than PCA, which is why the DCS-derived basis weights can carry mechanistic meaning.","core_discovery":"The paper's own claim is that the structure-property relationships in aqueous-dispersion CNT fibers can be read out of a multimodal dataset by decomposing aggregation-size distribution data with NMF into seven interpretable basis distributions, then applying SHAP to a Random Forest trained on those basis weights plus effective CNT length, Raman I_G/I_D, density, linear density, and processing conditions. On the 40 measured fibers, fracture strength is controlled by small, uniformly distributed aggregates (DCS Basis000 and Basis001) and a small amount of medium-sized aggregates (Basis004), while even a slight presence of large aggregates (Basis006) sharply lowers strength; electrical conductivity requires long effective CNT length (with a threshold near 2000 nm) and a narrow distribution of small aggregates; and Young's modulus is governed by macroscale density and fiber morphology rather than dispersion features. The authors interpret this as evidence that aggregate assemblies, not individual CNTs, are the fundamental structural unit determining CNT fiber properties.","pith_inferences":["A direct testable extension would be to check the stability of the seven NMF bases under bootstrap resampling of the 40 samples; if the bases change, the physical interpretation of Basis000/001/004/006 would need re-basing on independent data.","The paper leaves implicit that the identified thresholds (effective length around 2000 nm, I_G/I_D around 30) are dataset-specific; testing whether they survive across different CNT synthesis routes or dispersants would establish their generality.","Because density is the dominant factor for Young's modulus while dispersion dominates strength, the results imply competing processing priorities: densification may tighten packing but could disturb the narrow aggregate distribution, so multi-objective optimization is needed."],"forward_implications":["Fracture strength can be pursued by engineering dispersions with a narrow size distribution centered on small aggregates and only a minor medium-sized component, because these features maximize aggregate-aggregate junctions for load transfer.","Electrical conductivity can be pursued by lengthening CNTs and keeping their effective length above roughly 2000 nm while minimizing the broadening component of the small-aggregate distribution.","Young's modulus optimization should concentrate on spinning conditions that raise fiber density and reduce void content, since dispersion-state features have minor influence on stiffness.","The same EMML workflow is transferable to other nanomaterial-derived materials where multiscale structures and distribution-type measurements obscure which processing step matters."],"supporting_citations":[{"why":"Provides the dataset of 40 CNT fibers and their measured properties used for all regressions.","marker":"[10]"},{"why":"Supplies the disc centrifuge method that yields the aggregation size distributions decomposed by NMF.","marker":"[33]"},{"why":"Supplies the far-infrared spectroscopy method for measuring effective CNT length, a key conductivity feature.","marker":"[34]"},{"why":"Provides the NMF algorithm used for factor extraction from distribution data.","marker":"[37]"},{"why":"Provides the SHAP method used for global and local contribution analysis.","marker":"[32]"},{"why":"Supplies the weakest-link model used to interpret fracture strength in terms of local structure around defects.","marker":"[39]"},{"why":"Prior multimodal machine learning approach for CNT composites that this work extends with explainability.","marker":"[28]"},{"why":"Prior evidence that dispersion state matters for spinning processes, supporting the importance of DCS features.","marker":"[38]"},{"why":"Establishes the Raman spectroscopy basis for interpreting the I_G/I_D ratio as a measure of graphitic crystallinity and defects.","marker":"[35]"}],"fun_headline_variants":["NMF+SHAP reveal CNT fiber: small aggregates boost strength, long tubes conductivity","Explainable ML: small uniform aggregates key to CNT fiber strength","Multimodal ML: long CNTs and small aggregates tune fiber properties","AI decodes CNT fiber: strength from small aggregates, conductivity from long tubes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central attribution depends on the seven-component NMF decomposition, chosen by maximizing leave-one-out prediction accuracy on the same 40 samples, being a stable and physically meaningful representation of the dispersion state rather than an overfit artifact.","fun_headline_variants_meta":{"raw":{"variants":["NMF+SHAP reveal CNT fiber: small aggregates boost strength, long tubes conductivity","Explainable ML: small uniform aggregates key to CNT fiber strength","Multimodal ML: long CNTs and small aggregates tune fiber properties","AI decodes CNT fiber: strength from small aggregates, conductivity from long tubes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000757,"raw_usage":{"total_tokens":3408,"prompt_tokens":1036,"completion_tokens":2372,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":652,"completion_tokens_details":{"reasoning_tokens":2288}},"tokens_in":652,"tokens_out":2372,"duration_ms":14631,"temperature":1.0,"reasoning_tokens":2288,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T12:52:43.853453+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Prepare fibers from dispersions whose DCS aggregate-size distributions are deliberately varied (for example, by controlled sonication or filtration) while holding other variables fixed, and check whether fracture strength and conductivity track the SHAP-identified basis weights; alternatively, bootstrap the 40-sample dataset to see whether the same seven basis components and their SHAP signs reappear.","supporting_citations":[{"cited_title":"Tajima et al., Carbon, 152 (2019) 1","cited_arxiv_id":null,"evidence_quote":"Provides the dataset of 40 CNT fibers and their measured properties used for all regressions."},{"cited_title":"Muroga et al., Advanced Science, 10 (2023) 2302508","cited_arxiv_id":null,"evidence_quote":"Supplies the disc centrifuge method that yields the aggregation size distributions decomposed by NMF."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the far-infrared spectroscopy method for measuring effective CNT length, a key conductivity feature."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the NMF algorithm used for factor extraction from distribution data."},{"cited_title":"Lin et al., ACS Nano, 17 (2023) 22821","cited_arxiv_id":null,"evidence_quote":"Provides the SHAP method used for global and local contribution analysis."},{"cited_title":"Nadler et al., Carbon, 46 (2008) 1384","cited_arxiv_id":null,"evidence_quote":"Supplies the weakest-link model used to interpret fracture strength in terms of local structure around defects."},{"cited_title":"Qin et al., Advanced Functional Materials, 20 (2020) 2003429","cited_arxiv_id":null,"evidence_quote":"Prior multimodal machine learning approach for CNT composites that this work extends with explainability."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the Raman spectroscopy basis for interpreting the I_G/I_D ratio as a measure of graphitic crystallinity and defects."}],"review_version":1}