{"id":"3afb9b19-f8a3-48d4-8930-78c869696186","arxiv_id":"2505.21428","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A DFT plus interpretable machine learning workflow predicts formate adsorption energies at metal-oxide interfaces with about 0.17 eV RMSE, identifying work function and oxygen content as key descriptors.","lead":"The authors built a dataset of 2,863 DFT-computed formate adsorption energies on indium and zinc oxide clusters supported on gold, copper, and palladium surfaces, and trained four machine learning models to predict them. Their best model reaches about 0.17 eV error on unseen sites, and the analysis points to the surface work function, linked to cluster oxygen content, as a key descriptor.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Clean-structure features may misrepresent relaxed adsorption sites; the paper's assertion that reconstruction is 'a lesser issue' lacks quantitative support and directly bounds the screening claim.","rationale":"The reader's conditional verdict is appropriate. Our stress test did not uncover a different load-bearing failure; the clean-to-relaxed transfer is the most consequential assumption and it is explicitly acknowledged but not quantified. The proposed test would either confirm the stated applicability or force a narrower claim. Other issues (single-split RMSE without uncertainty, withheld data/code, Pd convergence bias) are secondary and do not change the verdict. The descriptor-interpretation claim (work function dominance) is correlational but not load-bearing for the prediction RMSE. The paper has independent support in its large dataset, standard split and CV procedure, consistent RMSEs across models, and a useful comparison to foundational pre-trained models, which strengthens confidence in the core prediction result conditional on the clean-to-relaxed assumption holding within the stated applicability range.","tokens_in":23038,"tokens_out":4748,"duration_ms":58813,"concrete_test":"Take the full dataset (2,863 relaxed adsorption structures and their corresponding clean clusters). For each pair, align the slab and compute the RMSD of cluster atoms (excluding formate and adsorbate H/C/O) between the clean structure and the relaxed adsorption structure; also compare the final binding motif to the enumerated site type from Section II B. Then re-compute Table III test RMSEs stratified into bins of cluster RMSD and by motif-change flag. If RMSE stays ≈0.17 eV in the high-RMSD/motif-change bins, the 'lesser issue' claim is supported. If RMSE rises substantially (e.g., >0.1 eV) in those bins, the clean-to-relaxed transfer is the limiting factor, the conclusion should be narrowed accordingly, and the paper should report this as a quantitative applicability range.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section II D states that only features from clean cluster structures are used, while the prediction targets are adsorption energies of relaxed cluster+formate configurations (Section II A, Eq. 4). The central claim that the best model predicts unseen formate binding configurations with RMSE ≈0.17 eV therefore depends on the assumption that adsorption does not substantially change the local binding environment. The authors themselves flag this at the end of Section III B: the models 'cannot be expected to work well' if the adsorbate dissociates or the surface reconstructs, and their nanosilicate study [49] showed such failures. However, for the present clusters they only assert that 'cluster reconstruction seems to be a lesser issue' without reporting any quantitative measure of reconstruction or any check of whether enumerated on-top/bridge assignments survive DFT relaxation. If a fraction of the 2,863 data points relaxed to a different site or pulled cluster atoms out of place, the features no longer describe the target, and the reported RMSE masks this mismatch; for new clusters where such reconstruction is more frequent, screening predictions would degrade. This is the load-bearing modeling choice.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents a supervised machine-learning workflow for predicting formate adsorption energies at the metal–oxide interface of inverse catalysts, specifically In_yO_x and Zn_yO_x clusters supported on Au(111), Cu(111), and Pd(111). The data set contains 2,863 DFT-relaxed adsorption energies derived from 25,117 enumerated binding configurations across 141 stoichiometries and 1,629 cluster structures. A key feature of the workflow is that all model features are computed from clean cluster structures only, while the prediction targets are relaxation energies of the cluster-plus-formate systems. Four models are compared: RBF-GPR, XGBoost, SISSO, and WWL-GPR. The best model, WWL-GPR, achieves a total test RMSE of 0.174 eV, which the authors argue is useful for screening. Learning-curve analyses show that roughly 100–280 training points from an unseen material are needed to approach full-model accuracy. Interpretability analysis identifies the work function, correlated with cluster oxygen content, as the dominant descriptor, along with Pauling electronegativity and d-band features. A supplementary comparison shows that the foundational pretrained models CHGNet and MACE-MP-0a perform substantially worse than the trained models.","tokens_in":23288,"tokens_out":6496,"duration_ms":74543,"significance":"If the clean-to-relaxed transfer assumption holds, the workflow has clear practical value: it promises adsorption-energy estimates for many binding sites at the cost of a single DFT calculation of the clean cluster, and it identifies physically plausible descriptors with experimental relevance. The study is carefully executed in several respects: the train/test split is stratified by material and site type, hyperparameters are selected by 5-fold cross-validation, per-material test metrics are reported, learning curves for unseen materials are included, and the comparison with foundational pretrained models is disclosed in the supplement. The main significance, however, is bounded by the unquantified assumption that adsorption does not substantially reconstruct the cluster or change the binding site, an assumption the paper itself flags in Section III B. The work-function/oxygen-content correlation is a credible and testable physical finding, but the strength of the screening claim depends on closing the clean-to-relaxed gap quantitatively.","major_comments":[{"comment":"Features are computed only from clean cluster structures (Section II D), while the targets are relaxed adsorption energies from Eq. (4). The authors acknowledge in Section III B that the models 'cannot be expected to work well' if the adsorbate dissociates or the surface reconstructs significantly, and they cite their nanosilicate study [49] as an example of such failures. They then state that for the present clusters 'cluster reconstruction seems to be a lesser issue,' but no quantitative evidence is provided. This is the load-bearing assumption for the central 0.17 eV RMSE claim. Please report, for the 2,863 relaxed data points, what fraction of the initially enumerated on-top/bridge assignments survive DFT relaxation; provide statistics on cluster-atom displacements or RMSD between clean and adsorbate-covered clusters; and show how model errors vary with the magnitude of reconstruction. Without such a check, the reported RMSE may partly reflect a mismatch between the structure the features describe and the structure the target energy describes, and the screening claim for new clusters is not yet supported.","section":"II D, III B"},{"comment":"All reported RMSEs come from a single stratified 80/20 split, and the differences among the best models are small: WWL-GPR gives a total test RMSE of 0.174 eV versus 0.192 eV for XGBoost and 0.206 eV for RBF-GPR in Table III. Without repeated splits, bootstrap resampling, or another uncertainty estimate, it is not possible to determine whether the model ordering is statistically meaningful or whether the 0.17 eV figure itself has a substantial sampling uncertainty. Because the paper selects the 'best' model and uses this RMSE as the headline screening metric, please provide confidence intervals or repeated-split statistics for the total and per-material RMSEs.","section":"III B, Table III"},{"comment":"The abstract states that the models can predict binding energies for 'unseen formate binding configurations' and can be used for prediction on 'structures outside of the original training data set.' The 0.17 eV RMSE, however, is for held-out configurations of the six material combinations already present in the training data. Section III B and Figure 4 show that for a completely unseen material the RMSE is on the order of 0.1 eV worse and that roughly 100–280 training points from that material are needed to approach full-model accuracy. Please qualify the screening claim in the abstract and conclusion so that it refers to unseen configurations and sites on materials represented in the training set, and state explicitly the additional data cost for genuinely new material combinations.","section":"Abstract, III B"}],"minor_comments":[{"comment":"The column headers in Table III are garbled in the manuscript text; please reformat the table so that each model has a clear, unambiguous column header.","section":"Table III"},{"comment":"The rendered axis labels and legend for Figure 5 are incomplete; please ensure that the feature names and the importance-score axis are fully visible and labeled.","section":"Figure 5"},{"comment":"The manuscript uses both 'Pauling electronegativity' (Section III C) and 'Pauli electronegativity' (Table S1); please standardize to 'Pauling electronegativity.'","section":"Table S1 and Section III C"},{"comment":"There is a duplicated article in Section I: 'where the the endothermic RWGS reaction'; please correct this typo.","section":"Introduction"},{"comment":"The paper notes that some slow-converging calculations, particularly on Pd-supported systems, were omitted; please state whether the omitted configurations are distributed uniformly across site types and stoichiometries, since a systematic omission could bias per-material error metrics.","section":"III A"},{"comment":"The rung-3 SISSO descriptor uses a primary-feature subset preselected by XGBoost feature importance, which is acknowledged in the supplement; please also state this caveat in the main-text interpretation paragraph, since it complicates the cross-model comparison of feature importance.","section":"III C and S3.1.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a competent and well-documented supervised learning study, and the central finding is potentially publishable once the clean-to-relaxed reconstruction assumption is quantified. The requested reconstruction analysis is feasible from existing DFT relaxation data and does not require new calculations. I see no grounds for rejection, but the current manuscript should not be accepted without this quantitative support."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth your attention. This is a competent, honest machine-learning study that targets a real bottleneck: DFT screening of inverse catalysts is expensive, and the authors show a practical workflow with a genuinely new dataset (formate on In/Zn oxide clusters on Au/Cu/Pd, 2,863 relaxed adsorption energies) and a clean-structure-only feature design. The best model (WWL-GPR) reaches an out-of-sample RMSE of 0.174 eV, and the learning curves indicate that a few hundred DFT points per new material get you close to that. The physical insight that the work function correlates with cluster oxygen content and drives adsorption trends is credible and useful for catalyst design. The FPM comparison in the supplement (CHGNet and MACE-MP-0a being clearly worse) is a nice addition and strengthens the case for the simple supervised models.\n\nThe main load-bearing assumption is that features from clean cluster structures can predict adsorption energies of relaxed cluster+formate configurations. The authors acknowledge at the end of Section III B that the models will fail if the adsorbate dissociates or the surface reconstructs significantly, and they cite their own nanosilicate work where exactly that happened. But for the present clusters they only assert that reconstruction is “a lesser issue” without showing quantitative evidence. The stress-test note is right: if a non-negligible fraction of the 2,863 relaxed structures moved to a different binding site or pulled cluster atoms out of place, the clean-structure features no longer describe the target, and the reported RMSE would be masking that mismatch. This is not a fatal flaw—the screening claim could still hold—but the paper should quantify it. A simple check, e.g., comparing site assignments before and after relaxation for a sample, or reporting the fraction of cases where the formate stayed at the enumerated site, would settle it.\n\nOther gaps are minor but real: model rankings are reported from a single test split without uncertainties across repeated splits, so the WWL-GPR vs XGBoost differences are not statistically established. The data and code are promised on Zenodo only upon publication, which limits independent verification now. The slow-convergence issue for Pd is disclosed but not analyzed for bias. None of these overturn the central RMSE claim, which is internally well-supported.\n\nThis paper deserves a serious referee. It is not field-reshaping, but it is a solid, reproducible-in-principle contribution to computational catalysis. I would send it to review with a request to address the clean-to-relaxed gap quantitatively and to release code/data. For my own work, I would cite the dataset and the work-function–oxygen-content trend if they publish.","headline":"Solid, well-executed ML screening study for formate on inverse catalysts; the clean-to-relaxed feature assumption is the main soft spot and needs quantitative support.","tokens_in":23854,"tokens_out":985,"would_cite":true,"duration_ms":12881,"reading_group":"yes","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 formate adsorption energies at the metal–oxide interface of inverse catalysts can be predicted with about 0.17 eV RMSE from features of the clean cluster alone, making low-cost computational screening of CO2…","keywords":["interpretable machine learning","adsorption energy prediction","inverse catalysts","CO2 hydrogenation","formate binding","work function descriptor","metal–oxide interface","DFT screening"],"falsifier":"Select the configurations from the dataset where the DFT relaxation moves the formate to a different binding site or changes the cluster geometry beyond a small threshold; if predictions on that subset show errors much larger than the overall 0.17 eV RMSE while rigid configurations stay accurate, the clean-to-relaxed feature mapping is the limiting assumption.","tokens_in":22840,"feed_emoji":"🧪","tokens_out":9324,"duration_ms":88166,"temperature":0.7,"pith_summary":"The paper tries to establish that machine learning can predict adsorption energies at the metal–oxide interface of inverse catalysts without relaxing every binding configuration. It builds a dataset of 2,863 formate binding energies from density functional theory across In$_{y}$O$_{x}$ and Zn$_{y}$O$_{x}$ clusters on Au(111), Cu(111), and Pd(111), using only features of the clean cluster as inputs. The best model, a graph-based Gaussian process, reaches an out-of-sample RMSE of about 0.17 eV, and the interpretable models indicate the work function — itself correlated with cluster oxygen content — is the dominant descriptor. If correct, this provides a cheap route to screening active sites and supplying adsorption energies for microkinetic models of CO2 hydrogenation.","feed_headline":"Adsorption energies predicted to 0.17 eV from clean clusters","feed_subtitle":"Workflow predicts formate binding across six inverse catalysts from one clean DFT structure, naming work function key.","key_machinery":"The load-bearing construct is the clean-to-relaxed mapping: every model is trained on features computed only from clean cluster structures — stoichiometric, geometric, and electronic descriptors including work function, Bader charges, and projected density-of-states moments — while targets are adsorption energies from fully relaxed cluster+formate configurations. Binding sites are enumerated as on-top and bridge sites, and a SOAP fingerprint with farthest-point sampling selects a diverse DFT subset. The interpretable SISSO descriptor is a linear expression of rung-limited algebraic features; its identified descriptors (Equations (6)–(9)) and XGBoost feature importance together carry the physical argument that work function, d-band center/width/filling, and electronegativity control adsorption.","core_discovery":"The central claim is that adsorption energies of formate at the metal–oxide interface can be predicted from a single DFT calculation of the clean cluster, without relaxing each adsorbate configuration. The authors enumerate on-top and bridge sites on In$_{y}$O$_{x}$ and Zn$_{y}$O$_{x}$ clusters on Au(111), Cu(111), and Pd(111), sample diverse configurations by farthest-point sampling in SOAP space, and train four models on 2,863 relaxed DFT energies. The WWL-GPR graph model achieves the lowest test RMSE (0.174 eV), followed by XGBoost (0.192 eV) and RBF-GPR (0.206 eV), while the interpretable SISSO descriptor performs worse (0.261 eV) but reveals structure. Feature importance and SISSO terms both put the work function first; work function correlates with cluster oxygen content, and formate binds more strongly on oxygen-poor clusters where metal atoms lack oxygen coordination.","pith_inferences":["Editorial extension: because work function is experimentally accessible, the descriptor ranking suggests a fast experimental prescreen by work function before DFT, a step the paper itself does not propose.","Editorial extension: the clean-to-relaxed assumption is likely to be tested harder by more reactive adsorbates such as CO or atomic H than by formate, since those bind more invasively and may induce reconstruction; a hybrid workflow that re-checks reconstructed sites with a surrogate potential would be a natural extension.","Editorial extension: the disagreement between XGBoost and SISSO over electronegativity's role suggests that descriptor importance is partly a function of model class; combining both views may be needed before drawing mechanistic conclusions, which the paper does not resolve."],"forward_implications":["A single DFT calculation of the clean cluster can yield screening-level adsorption energies (RMSE about 0.17 eV) for all enumerated sites, letting expensive DFT relaxations be reserved for promising candidates.","Introducing about 280 DFT data points from an unseen material brings test RMSE within 0.02 eV of the full-data model, so extending the workflow to new metal–oxide combinations requires only a moderate data investment.","Cluster oxygen content, through its correlation with the work function, is a practical tuning knob: oxygen-poor clusters bind formate more strongly, a trend directly relevant to CO2 hydrogenation activity.","The models cannot be expected to work when the adsorbate dissociates or the surface reconstructs strongly; within the six studied systems such restructuring is reported to be a lesser issue.","The workflow can likely be extended to other adsorbates and binding motifs, allowing simultaneous prediction of many site/adsorbate combinations from the same clean-cluster calculation."],"supporting_citations":[{"why":"Supplies the clean inverse-catalyst structures and stable stoichiometries from the prior global-structure-optimization and ab initio thermodynamics study, defining the dataset space.","marker":"[25]"},{"why":"Provides the SISSO compressed-sensing method that constructs the interpretable analytical descriptors used for physical insight.","marker":"[44]"},{"why":"Introduces the WWL graph kernel and GPR model that yields the most accurate adsorption-energy predictions in the paper.","marker":"[45]"},{"why":"Provides the XGBoost algorithm used for high-accuracy predictions and feature-importance analysis.","marker":"[43]"},{"why":"Defines the SOAP descriptor used to fingerprint binding configurations for diversity sampling before DFT relaxation.","marker":"[26]"},{"why":"Supplies the scikit-learn implementations used for precomputed distance matrices, RBF-GPR, and model-selection routines.","marker":"[29]"},{"why":"Formulates the d-band model linking electronic structure of the binding site to adsorption strength, used to explain feature importance.","marker":"[52]"},{"why":"Demonstrates the failure mode for amorphous nanosilicates where adsorption restructures the cluster, marking the boundary of the clean-to-relaxed assumption.","marker":"[49]"},{"why":"Pre-trained CHGNet interatomic potential used as a baseline to benchmark the workflow's prediction errors in the supplementary material.","marker":"[47]"},{"why":"Pre-trained MACE-MP-0a interatomic potential used as a second baseline in the supplementary comparison.","marker":"[48]"}],"fun_headline_variants":["Work function drives adsorption at metal-oxide interfaces","ML predicts adsorption from clean clusters, no relaxations","One clean structure forecasts CO2 hydrogenation binding","Interpretable ML reveals work function controls binding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that features of the clean cluster alone carry enough information to predict the adsorption energy of the relaxed cluster-plus-formate configuration; this fails if adsorption dissociates the formate or significantly reconstructs the cluster, a limitation the authors state at the end of Section III B and connect to a prior nanosilicate study.","fun_headline_variants_meta":{"raw":{"variants":["Work function drives adsorption at metal-oxide interfaces","ML predicts adsorption from clean clusters, no relaxations","One clean structure forecasts CO2 hydrogenation binding","Interpretable ML reveals work function controls binding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000182,"raw_usage":{"total_tokens":1323,"prompt_tokens":971,"completion_tokens":352,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":293}},"tokens_in":587,"tokens_out":352,"duration_ms":4336,"temperature":1.0,"reasoning_tokens":293,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:28:44.962838+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Select the configurations from the dataset where the DFT relaxation moves the formate to a different binding site or changes the cluster geometry beyond a small threshold; if predictions on that subset show errors much larger than the overall 0.17 eV RMSE while rigid configurations stay accurate, the clean-to-relaxed feature mapping is the limiting assumption.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the clean inverse-catalyst structures and stable stoichiometries from the prior global-structure-optimization and ab initio thermodynamics study, defining the dataset space."},{"cited_title":"Xu , author K","cited_arxiv_id":null,"evidence_quote":"Introduces the WWL graph kernel and GPR model that yields the most accurate adsorption-energy predictions in the paper."},{"cited_title":"Pedregosa , author G","cited_arxiv_id":null,"evidence_quote":"Supplies the scikit-learn implementations used for precomputed distance matrices, RBF-GPR, and model-selection routines."},{"cited_title":"Hammer \\ and\\ author J","cited_arxiv_id":null,"evidence_quote":"Formulates the d-band model linking electronic structure of the binding site to adsorption strength, used to explain feature importance."},{"cited_title":"Andersen \\ and\\ author A","cited_arxiv_id":null,"evidence_quote":"Demonstrates the failure mode for amorphous nanosilicates where adsorption restructures the cluster, marking the boundary of the clean-to-relaxed assumption."}],"review_version":1}