{"id":"fa47fe77-284a-4705-a4ed-f71793227e43","arxiv_id":"2501.08919","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A multimodal machine-learning framework trained on simulated XAS/EELS spectra infers local element counts and oxygen-vacancy/antisite defects in NMC battery cathode materials.","lead":"This paper trains machine learning on simulated X-ray and electron spectroscopy data to infer local atomic arrangements in battery cathode materials. The method combines multiple spectral channels, oxygen and transition-metal edges, and is tested on experimental lithium-ion battery data with claims of improved accuracy over single-spectrum approaches.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Antisite defect detection may be an artifact of mismatched DFT functionals: antisite structures use PBE+U while pristine use SCAN+U, so the 100% classification accuracy may reflect theory level, not physical defects.","rationale":"The reader's weakest assumption concerns the simulation-to-experiment transfer gap. My analysis identifies a more internal and more immediate threat: the simulated defect and pristine spectra are not generated on equal footing. The functional/U mismatch for antisite structures means the classifier can exploit a trivial theory-level signature. This is a concrete, testable flaw that directly affects the headline claim about defect detection. The oxygen-vacancy classification does not suffer from this functional mismatch, and the Li-content inference with RMSE < 0.1 against capacity-derived values is a genuine experimental validation. However, the antisite result is central to the abstract, and the current manuscript provides no control. A conditional acceptance requiring the matched-functional recalculation is the appropriate outcome; if the accuracy collapses, the claim should be substantially revised. The reader's verdict is already CONDITIONAL, so my concern does not change the verdict category but adds a specific technical condition that must be met.","tokens_in":14862,"tokens_out":3170,"duration_ms":34385,"concrete_test":"Recompute the 14–18 antisite-defective structures with the same SCAN+U settings and Hubbard U values used for the pristine structures, regenerate the FDMNES spectra, and retrain/evaluate the antisite-vs-pristine classifier. If the accuracy drops materially below 100% or the most important spectral features change, the current result is an artifact of the PBE-vs-SCAN difference. As a cheaper control, train a binary classifier to distinguish PBE+U from SCAN+U spectra ignoring defect labels; high accuracy would confirm the confound.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the multimodal model detects antisite defects (Figure 4d, 100% accuracy) is undermined by a simulation-level confound. In the 'DFT and Spectra Simulations' section, pristine and oxygen-vacancy structures are computed with the SCAN functional and Hubbard U values of 2.43 eV (Ni), 2.93 eV (Mn), and 2.86 eV (Co). For antisite defects, the text states: 'the exchange-correlation was treated using the Perdew–Burke–Ernzerhof (PBE) generalized gradient approximation (GGA)... Correspondingly, Hubbard U values were adjusted to 6.7 eV for Ni, 4.2 eV for Mn, and 4.9 eV for Co.' The FDMNES core-level spectra are generated from these relaxed structures, so the antisite-defective spectra carry a systematic PBE+U signature that is absent from the pristine SCAN+U spectra. A classifier can therefore separate the two classes by recognizing the DFT functional rather than the defect. This concern is sharpened by the suspiciously perfect 100% accuracy on a very small sample: the text reports 14 antisite structures in 'Structures and Spectra' but 18 in the 'Local defect' subsection. No control or ablation for the functional mismatch is reported, and there is no experimental ground truth for antisite defects. Consequently, the abstract's assertion that the model can determine whether antisites are present is not currently supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a machine-learning workflow that fuses multimodal core-level spectroscopy (O K-edge and Ni/Mn/Co L2,3-edges) with ab initio simulations to infer local structure and defects in NMC cathode materials. The authors train XGBoost models on FDMNES-simulated spectra of pristine, oxygen-vacancy, and Ni/Li antisite structures, using CDF featurization and Bayesian hyperparameter optimization. They report accurate Li-content inference on experimental sXAS data, improved local-composition inference over unimodal approaches, and 100% classification accuracy for oxygen vacancies and antisite defects. The paper also includes a permutation-importance analysis and a qualitative validation on Si-doped Li-rich NMC via EELS mapping.","tokens_in":15184,"tokens_out":3212,"duration_ms":36843,"significance":"If the main claims hold, the work would be a valuable demonstration that combining multiple spectroscopic edges improves ML-based inference of local composition and defects, with a practical pathway from simulations to experimental EELS/XAS data. The manuscript has several concrete strengths: a systematic benchmarking of five spectrum simulation codes against experimental spectra, a thoughtful comparison of featurization and ML models, and a clear attempt to test the trained model on experimental data and on a doped material system. The interpretability analysis (feature importance mapped to energy domain) also adds physical insight. However, the defect-detection claims, especially for antisites, are currently under-supported because of a simulation-level confound and very small, inconsistent dataset sizes. The experimental Li-content validation also lacks error bars and calibration details. The central claim that the approach detects defects that are 'impossible' for other techniques is overstated relative to the evidence presented.","major_comments":[{"comment":"The antisite-defect classification is undermined by a functional mismatch between the training classes. Pristine and oxygen-vacancy structures are relaxed with SCAN+U (U_Ni=2.43 eV, U_Mn=2.93 eV, U_Co=2.86 eV), while antisite structures use PBE+U with much larger U values (6.7, 4.2, 4.9 eV). The FDMNES core-level spectra inherit these differences, so a classifier can separate the classes by recognizing the exchange-correlation functional rather than the physical defect. This is a load-bearing issue for the abstract's claim that the model 'determine[s] whether local defects such as ... antisites are present.' The authors should provide a control: for example, compute a set of pristine (or oxygen-vacancy) structures with the same PBE+U settings as the antisite structures and show that the classifier still distinguishes antisites from these matched controls. Without such a control, the 100% accuracy in Figure 4(d) cannot be attributed to the antisite defect itself.","section":"Methods, DFT and Spectra Simulations; Local defect; Figure 4(d)"},{"comment":"The defect classification results are reported as exact '100% accuracy' on datasets that are very small and are not described with a clear evaluation protocol. The text gives 14 antisite structures in 'Structures and Spectra' but 18 in the 'Local defect' subsection, and it does not state how many structures were used for training versus testing, whether repeated train/test splits or cross-validation were performed, or how class weighting affected the reported accuracy. The oxygen-vacancy result is similarly based on 136 defective structures against 701 pristine, with no error bars or confidence intervals. The authors should provide the full classification details: the number of test instances, the train/test split method (e.g., leave-one-out or stratified k-fold), the distribution of predicted probabilities, and confusion matrices. I also recommend reporting precision/recall rather than only a single accuracy number, since the class imbalance and the very small antisite count could make the 100% value an artifact of the split.","section":"Structures and Spectra; Local defect; Figure 4(c,d)"},{"comment":"The experimental Li-content validation lacks sufficient quantitative detail. The text reports 'RMSE value of less than 0.1' but does not specify the number of data points, the composition/cycle states covered, or the uncertainty in either the predicted or the capacity-derived Li content. The capacity-derived 'ground truth' relies on the assumptions in the equation (e.g., one electron per Li, no side reactions), and the manual shift/broadening calibration applied to simulated data is described only qualitatively. To support the claim of quantitative agreement, the authors should list the individual Li% values with uncertainties for the points in Figure 3, state how the calibration parameters were chosen (e.g., grid search or fixed offsets), and ideally report the per-condition errors. The stronger claim that the method yields local Li content from EELS/XAS mapping is not demonstrated here, since the comparison in Figure 3 is only against bulk-averaged capacity-derived values; a validation on spatially resolved experimental data would be needed to support that extension.","section":"Li Content Inference; Figure 3"},{"comment":"The local environment inference uses a hand-set placeholder for absent elements: a constant line at y=0.1 with 2% Poisson noise. The choice of 0.1 is not justified, and no sensitivity analysis is provided for this value. Since the absent-element spectra constitute part of the input features for predicting the counts of Mn and Co, the reported accuracy gains of the multimodal method over unimodal ones may depend on this arbitrary placeholder. The authors should vary this placeholder (e.g., 0.0, 0.01, 0.5) and show that the qualitative conclusion of multimodal superiority is unchanged, or replace the placeholder with a physically motivated baseline (e.g., a flat pre-edge spectrum from a dilute reference).","section":"Local Environment Inference; Figure 4(a,b)"}],"minor_comments":[{"comment":"The term 'RSME' is a typo; it should be 'RMSE' (root mean square error). Please also check for other typos throughout the manuscript, including the duplicated paragraph about BO-TPE that appears twice in the 'Machine Learning Model' and 'Model Construction and Training' sections.","section":"Li Content Inference"},{"comment":"The figure and text contrast multimodal (red) and unimodal (black/gray) workflows, but the caption and text do not define which edges are used in each unimodal baseline (e.g., only O K-edge vs. only Ni L-edge). Please specify the exact unimodal feature sets for all comparisons in Figures 3 and 4.","section":"Figure 1 and workflow description"},{"comment":"No data or code availability statement is included. I would encourage the authors to share the training data (simulated spectra and structure labels) and the trained models, as this would strengthen reproducibility and allow the community to build on the workflow.","section":"Data and code availability"},{"comment":"The feature importance plot is mapped back to the energy domain, but the text only qualitatively discusses the Ni L3 and O K-edge peaks. Please state the exact energy ranges highlighted in the plot and, if possible, provide a quantitative measure (e.g., integrated importance over a window) to support the claim that Ni L3 is 'the most important feature.'","section":"Model Interpretation, Figure 6"},{"comment":"The claim that defect detection is 'impossible for single mode spectra or other experimental techniques' is too strong and is not evidenced by the data presented. The paper only compares to two unimodal ML baselines (O K-edge and one representative TM L-edge) and does not compare to, for example, combined O K + one TM edge or to non-ML experimental analysis methods. Please soften the wording or provide a more systematic baseline comparison.","section":"Local defect subsection"}],"recommendation":"major_revision","confidential_remarks":"The core idea of the paper is timely and the simulation-to-experiment workflow is potentially valuable to the community. However, the antisite functional mismatch is a serious confound that must be addressed before the defect-detection claim can be accepted. In addition, the experimental validation and evaluation protocols need to be tightened. If the authors can provide the requested control calculations and detailed error analysis, the revised manuscript would be a strong contribution. I would also suggest the editor ask the authors to clarify the discrepancy in the number of antisite structures."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The strong part is Li content inference: XGBoost on FDMNES-simulated O K and TM L2,3 edges, applied to experimental soft XAS of cycled NMC, gives RMSE < 0.1 against capacity-derived Li content. Multimodal beats single-edge by 30-50%. That is credible and useful. The weak part is the antisite defect claim. Antisite structures were relaxed with PBE+U (U = 6.7/4.2/4.9 for Ni/Mn/Co) while all pristine and oxygen-vacancy structures use SCAN+U (2.43/2.93/2.86). Since FDMNES spectra inherit the electronic structure of these calculations, the classifier can separate antisite from pristine by detecting the DFT functional, not the defect. There is no control or ablation, no experimental ground truth, and the 100% accuracy is on 14-18 structures. That claim is unsupported.\n\nWhat is genuinely new: multimodal fusion of O K and TM L edges for local composition/defect classification in NMC is not in the prior single-edge ML literature. The featurization and model benchmarking is careful. The Li content result is the real contribution.\n\nSoft spots: data and code are absent. The manual shift/broadening calibration is underreported — how were these chosen, and are they held fixed across samples? There is an inconsistency between 14 antisite structures in the Structures section and 18 in the Local defect section. The O-vacancy transfer test on Si-doped Li-rich NMC is promising but circumstantial: overlap between predicted vacancies and Si pixels does not prove the vacancies are there. The local-environment inference accuracy numbers are all on simulated data, so they establish only that the model can fit the simulation.\n\nWho this is for: anyone working on ML analysis of core-loss spectra, especially battery cathode degradation studies. It deserves a serious referee, but the antisite claim needs either a consistent-functional recomputation or an explicit ablation showing the classifier cannot separate PBE+U from SCAN+U. I would recommend major revision rather than acceptance. The Li content part can stand, but the headline needs to change.","headline":"Useful Li-content inference from multimodal spectra, but the antisite defect claim is likely an artifact of a DFT functional mismatch and should not be accepted without reanalysis.","tokens_in":15714,"tokens_out":3357,"would_cite":false,"duration_ms":35027,"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 machine-learning model that fuses oxygen and transition-metal core-level spectra can infer local composition and detect defects in NMC battery cathodes, a capability single-edge spectra do not provide.","keywords":["machine learning","multimodal spectroscopy","core-level spectroscopy","X-ray absorption spectroscopy","electron energy loss spectroscopy","NMC cathode materials","oxygen vacancies","antisite defects"],"falsifier":"A decisive check would compare the model's per-pixel defect predictions with independently determined defect maps on the same particles: for example, oxygen-vacancy predictions from the multimodal model versus atomically resolved STEM imaging or quantified EELS reference standards on Si-doped Li-rich NMC. If the model flags vacancies where the independent structure shows a stoichiometric lattice, or misses them where the lattice clearly has vacancies, the simulation-to-experiment transfer premise collapses.","tokens_in":14690,"feed_emoji":"🔋","tokens_out":13735,"duration_ms":129740,"temperature":0.7,"pith_summary":"This paper sets out to show that a machine-learning model fed several core-level spectra at once—the oxygen K-edge plus the nickel, manganese, and cobalt L2,3 edges—can infer local atomic structure that no single spectral stream exposes. The test system is lithium nickel manganese cobalt oxide (NMC), the battery-cathode family, studied across lithiation states, oxygen vacancies, and Ni/Li antisite defects. Training uses spectra simulated from first-principles structures, and the model then transfers to measured X-ray absorption and electron energy-loss data. If the transfer is reliable, the approach turns ordinary spectroscopy maps into maps of local composition and defect content, and the same recipe could be retrained for other multicomponent materials.","feed_headline":"Fused spectral edges let AI spot defects single edges miss","feed_subtitle":"Trained on computed spectra, the model transfers to real X-ray and electron energy-loss maps of battery cathodes","key_machinery":"The central object is the multimodal spectrum vector: for each local site, the O K-edge and the Ni, Mn, and Co L2,3-edge spectra, converted into cumulative distribution functions (CDFs) to reduce sensitivity to energy-axis shifts and fed together to gradient-boosted decision trees. Training spectra are computed from first-principles structures of 851 NMC configurations, including pristine, oxygen-vacant, and Ni/Li antisite structures, using a spectral simulation method selected by comparison with experimental reference spectra; the simulated curves are then aligned to experiment by manual energy shifting and broadening. The fusion itself carries the argument: defect classification reaches 100% only when all four edges are supplied, and feature-importance maps assign physical meaning to the channels, with the Ni L3 edge dominating Li-content prediction and the O K-edge reporting oxygen–transition-metal bonding and charge compensation.","core_discovery":"The central claim is that fusing element-specific spectral edges changes what spectroscopy can say about a material: a gradient-boosted tree trained on the O K-edge together with the Ni, Mn, and Co L2,3 edges classifies oxygen vacancies and Ni/Li antisite defects with perfect accuracy on simulated test cases, while single-edge models miss them. The same multimodal model predicts local Li, Ni, Mn, and Co counts inside a roughly 0.3 nm coordination shell, and it infers bulk Li content from experimental soft X-ray spectra of cycled NMC cathodes with root mean square error below 0.1, beating single-edge inference by 30–50%. When applied without retraining to a Si-doped Li-rich NMC sample, the model's predicted oxygen-vacancy pixels overlap with the measured Si L-edge signal, matching a prior experimental finding that silicon promotes oxygen vacancies. The paper reads this as evidence that multimodal core-level spectroscopy, after calibrating simulated spectra to experimental energy scales, can reveal local defects that a single spectrum or other experimental techniques cannot.","pith_inferences":["The 100% defect-classification accuracies are simulated-test-set numbers; real experimental ground truth for defects is harder to certify, so the transfer to arbitrary samples is the least-tested part of the pipeline, with the Si-doped co-localization serving as an indirect check.","Because CDF inputs are robust to energy-axis shifts, the framework may transfer across instruments with different energy calibrations with little or no recalibration, an extension the paper does not test.","The same training recipe could be adapted to neighboring layered-oxide chemistries, such as sodium analogues or other dopant variants, using a small transfer-learning dataset as the paper lists for future work.","The local-environment definition of a 0.3 nm sphere ties predictions to a specific spatial scale; probing atomic-column resolution would reveal whether the multimodal advantage persists when the coordination shell contains far fewer atoms."],"forward_implications":["Local Li content can be mapped directly from XAS or EELS spectrum images, with a reported root mean square error under 0.1 for cycled NMC cathodes, instead of relying on bulk capacity averages.","Oxygen vacancies and Ni/Li antisites become detectable from core-level spectra when several edges are fused; in this study single-edge models cannot achieve that detection.","A model trained on simulated spectra can be applied without retraining to a chemically related but different system, Si-doped Li-rich NMC, and its defect predictions agree with an independent experimental report.","Feature-importance analysis yields physically interpretable channels: the Ni L3 edge tracks nickel redox and lithium content, while the two O K-edge features report oxygen–transition-metal bonding and charge compensation."],"supporting_citations":[{"why":"Supplies the first-principles relaxed NMC structures whose spectra form the training set.","marker":"47"},{"why":"Computes the oxygen and transition-metal core-level spectra used as machine-learning inputs.","marker":"62"},{"why":"Introduces the cumulative-distribution-function featurization adopted for its shift robustness.","marker":"24"},{"why":"Establishes the tree-based modeling precedent for X-ray absorption near-edge spectra that this work extends to multimodal inputs.","marker":"69"},{"why":"Provides experimental NMC reference spectra used to benchmark the simulation methods.","marker":"73"},{"why":"Supplies the soft X-ray absorption measurements on NMC622, NMC721, and NMC811 used to test Li-content prediction.","marker":"74"},{"why":"Reports the Si-promoted oxygen-vacancy behavior used to validate the model on Si-doped Li-rich NMC.","marker":"44"},{"why":"Describes the synthesis of the Si-doped Li-rich NMC sample measured by EELS.","marker":"75"},{"why":"Provides the antisite-defect supercell construction used to generate training structures.","marker":"59"},{"why":"Supports the choice of oxygen-vacancy structures at low lithium content.","marker":"58"}],"fun_headline_variants":["AI fuses spectral edges to reveal hidden material defects","Multimodal spectra plus ML uncovers battery cathode defects","Fused X-ray and electron spectra expose local defects","AI on fused spectral edges detects defects single modes miss","Machine learning reads multiple spectral edges to spot atomic defects"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that simulated core-level spectra, after manually chosen energy shifts and broadening, are faithful stand-ins for real experimental spectra for every composition, lithiation state, and defect type used here.","fun_headline_variants_meta":{"raw":{"variants":["AI fuses spectral edges to reveal hidden material defects","Multimodal spectra plus ML uncovers battery cathode defects","Fused X-ray and electron spectra expose local defects","AI on fused spectral edges detects defects single modes miss","Machine learning reads multiple spectral edges to spot atomic defects"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000579,"raw_usage":{"total_tokens":2774,"prompt_tokens":1039,"completion_tokens":1735,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":655,"completion_tokens_details":{"reasoning_tokens":1658}},"tokens_in":655,"tokens_out":1735,"duration_ms":11448,"temperature":1.0,"reasoning_tokens":1658,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:14:44.340034+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check would compare the model's per-pixel defect predictions with independently determined defect maps on the same particles: for example, oxygen-vacancy predictions from the multimodal model versus atomically resolved STEM imaging or quantified EELS reference standards on Si-doped Li-rich NMC. If the model flags vacancies where the independent structure shows a stoichiometric lattice, or misses them where the lattice clearly has vacancies, the simulation-to-experiment transfer premise collapses.","supporting_citations":[{"cited_title":"Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set","cited_arxiv_id":null,"evidence_quote":"Supplies the first-principles relaxed NMC structures whose spectra form the training set."},{"cited_title":"Self-consistent aspects of x-ray absorption calculations","cited_arxiv_id":null,"evidence_quote":"Computes the oxygen and transition-metal core-level spectra used as machine-learning inputs."},{"cited_title":"J.; Yang, W.; Sun, C.; Lee, G.; McReynolds, D.; Allan, D.; Marulanda Arias, J","cited_arxiv_id":null,"evidence_quote":"Introduces the cumulative-distribution-function featurization adopted for its shift robustness."},{"cited_title":"B.; Carbone, M","cited_arxiv_id":null,"evidence_quote":"Establishes the tree-based modeling precedent for X-ray absorption near-edge spectra that this work extends to multimodal inputs."},{"cited_title":"Atomic resolution structural and chemical imaging revealing the sequential migration of Ni, Co, and Mn upon the battery cycling of layered cathode","cited_arxiv_id":null,"evidence_quote":"Provides experimental NMC reference spectra used to benchmark the simulation methods."},{"cited_title":"Oxygen Activities Governing Structural Reversibility in Industrial Ni-Rich Layered Cathodes","cited_arxiv_id":null,"evidence_quote":"Supplies the soft X-ray absorption measurements on NMC622, NMC721, and NMC811 used to test Li-content prediction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Reports the Si-promoted oxygen-vacancy behavior used to validate the model on Si-doped Li-rich NMC."},{"cited_title":"R.; Sheldon, B","cited_arxiv_id":null,"evidence_quote":"Describes the synthesis of the Si-doped Li-rich NMC sample measured by EELS."},{"cited_title":"Role of Superexchange Interaction on Tuning of Ni/Li Disordering in Layered Li(Ni x Mn y Co z )O2","cited_arxiv_id":null,"evidence_quote":"Provides the antisite-defect supercell construction used to generate training structures."},{"cited_title":"Tuning of Thermal Stability in Layered Li(NixMnyCoz)O2","cited_arxiv_id":null,"evidence_quote":"Supports the choice of oxygen-vacancy structures at low lithium content."}],"review_version":1}