{"id":"578a09c9-4eef-4762-8088-ecd2963a844f","arxiv_id":"2607.17463","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"On 47 intermetallic alloys, s- and p-band descriptors dominate d-band center in predicting N2/N2H/NH3 adsorption, with Fe9Co7 and Fe3Co as the most balanced NRR candidates.","lead":"Using DFT and machine learning on about 1,200 adsorption sites across 47 two-metal alloys, this study reports that s- and p-electron descriptors predict nitrogen-reduction binding better than the traditional d-band center, and singles out Fe9Co7 and Fe3Co as balanced catalysts for ammonia synthesis.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Global SHAP 's/p over d' conclusion may be driven by the Al/Zn/Cu/Pt majority, not by partially filled d-band TMs; Fe-Co PDOS evidence is anecdotal.","rationale":"The reader's CONDITIONAL verdict already captures the need for further validation; my concern refines the specific condition. The DFT convergence issue (1x1x1 k-points, unspecified spin polarization) is a real risk to the underlying data, but the more directly load-bearing problem for the paper's central claim is the composition confounding in the SHAP analysis. The authors acknowledge the majority of configurations contain Al/Zn/Cu/Pt, and their global feature-importance conclusion could be an expected consequence of that composition distribution rather than a discovery about Fe/Co/Ni/V systems. The PDOS analysis of Fe3Co and Fe9Co7 is suggestive but not quantitative, and the absence of a d-band-only ablation means the 'beyond d-band' assertion is not rigorously benchmarked. These concerns do not invalidate the dataset or the ML models, but they do mean the headline claim should be presented as conditional on subset-specific analysis. The reader's verdict of CONDITIONAL remains appropriate; no change is needed, but the condition should explicitly include a composition-stratified SHAP check and a d-band-only baseline.","tokens_in":23728,"tokens_out":2958,"duration_ms":29828,"concrete_test":"Retrain XGB and compute SHAP on the subset of adsorption configurations whose active-site and neighbor atoms are exclusively Fe, Co, Ni, and V (excluding all Al/Zn/Cu/Pt). Also train a d-band-only model (εd, fd only) and compare test MAE/SHAP to the full 20-feature model. If s/p features do not outrank d-band features in the TM-only subset, or if the d-band-only baseline matches the full model's accuracy, the central 'beyond d-band' claim is unsupported for partially filled d-band intermetallics.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that s- and p-band characteristics are more important than the d-band center for *N2 and *N2H on intermetallic NRR catalysts, including transition metals with partially filled d-bands—rests on SHAP feature rankings from ML models trained on the full dataset (Section 3.2). However, the authors themselves note that ~800 of ~1200 configurations contain Al, Zn, Cu, or Pt, i.e., p-block or filled/nearly-filled d-band elements where s/p importance is expected. The global SHAP ranking therefore cannot distinguish whether s/p dominance is intrinsic to partially-filled d-band TM systems or an artifact of dataset composition. The only evidence offered for the TM-specific claim is the PDOS analysis of Fe3Co and Fe9Co7 (Fig. 9), which is qualitative and limited to two surfaces. Moreover, no d-band-only baseline was trained; the 'more prominent' claim is not quantified as an improvement over a model using only d-band descriptors. Thus the paper's headline 'beyond the d-band center' is not directly demonstrated for the regime it claims to extend, and the SHAP-based support is confounded by the very composition classes where s/p dominance is already known. This is load-bearing because the strongest claim and the catalyst recommendation both depend on this generalization.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a high-throughput DFT and machine-learning study of NRR-relevant adsorption energies (*N2, *N2H, *NH3) on 47 bimetallic intermetallic compounds, totaling about 1,200 adsorption configurations. The authors compute adsorption energies with PBE, construct a 72-dimensional descriptor set combining intrinsic atomic properties, shell statistics, and s/p/d band centers and fillings, and train multiple ML models; XGBoost yields test MAEs of 0.26 eV (*N2), 0.39 eV (*N2H), and 0.17 eV (*NH3). SHAP analysis identifies max(covalent radius), s- and p-band descriptors as more important than the d-band center, and the authors interpret this as evidence of s/p–d hybridization that extends beyond p-block and filled-d-band systems. They further analyze scaling relations, find that *N2H–*NH3 scaling is broken, and propose Fe9Co7 and Fe3Co as the most balanced catalysts. The paper includes open code on GitHub and a detailed supporting information section.","tokens_in":24061,"tokens_out":3221,"duration_ms":30105,"significance":"If the DFT energies and the descriptor interpretation are reliable, this work would be a useful contribution to NRR catalyst screening by broadening the descriptor space beyond the d-band center and proposing specific non-noble intermetallic candidates. The manuscript's strengths include a comparatively exhaustive enumeration of adsorption sites, a transparent ML pipeline with a sealed test set, open data/code availability, and a physically interpretable feature set. However, the central 'beyond the d-band center' claim is currently supported by a global SHAP analysis that is confounded by dataset composition, and the quantitative DFT basis is not sufficiently validated for magnetic transition-metal surfaces. These issues are load-bearing for both the headline mechanistic claim and the catalyst ranking.","major_comments":[{"comment":"The DFT protocol is not sufficiently validated for the systems studied. Bulk and surface optimization uses a 1×1×1 Monkhorst–Pack k-point grid, and no slab thickness, vacuum gap, or spin-polarization settings are reported for Fe-, Co-, and Ni-containing surfaces. For magnetic transition metals, adsorption energies can shift by several tenths of an eV with k-point sampling and spin treatment; this uncertainty directly affects the Fe–Co ranking in Fig. 4 and the electronic-descriptor trends in Figs. 6 and 8. The authors should provide convergence tests in k-points, slab thickness, and vacuum, specify spin treatment, and quantify the resulting error bars on the adsorption energies.","section":"Section 2.1"},{"comment":"The SHAP-based conclusion that s- and p-band characteristics are 'more prominent' than the d-band center for *N2 and *N2H is confounded by dataset composition. The authors themselves note that ~800 of ~1200 configurations contain Al, Zn, Cu, or Pt—elements for which s/p dominance is expected from prior work. The global SHAP ranking therefore does not establish that s/p features are decisive for partially-filled d-band transition metals. The only evidence offered for the Fe–Co case is the qualitative PDOS analysis of Fig. 9, which covers two surfaces and is not quantified. Additionally, no d-band-only baseline model was trained, so the 'more prominent' claim is not measured as an improvement over a model using only d-band descriptors. Please add a d-band-only baseline and a subset analysis restricted to Fe/Co/Ni/V systems, or temper the claim accordingly.","section":"Section 3.2"},{"comment":"Outlier removal is performed on the full dataset before the train/test split: 'extreme outliers... are identified as samples beyond three standard deviations from the mean (|z|>3) of the target distribution' and 'were removed from the dataset for each adsorbate.' Since the mean and standard deviation are computed from the entire dataset, the test set influences which samples are removed, which is a data-leakage mechanism. This can inflate the reported test MAE/R². The outlier removal threshold and any filtering should be fit only on the training folds, or the authors should demonstrate that the model conclusions are unchanged when outliers are retained.","section":"Section 2.3"}],"minor_comments":[{"comment":"In Fig. 4(a), 'Al3N2' appears to be a typo for 'Al3Ni2'. Please check the label.","section":"Fig. 4"},{"comment":"The notation 'εb,local' and 'fb,local' is used in the caption but not defined there; define these symbols explicitly, including the energy window (-10 to 10 eV) used in Eq. (2)-(3).","section":"Fig. 6"},{"comment":"The 'asymmetry index' is introduced as the difference between maximum and minimum electronegativity of the four nearest neighbors, but it is not fully defined: is it the spread of electronegativities or the difference between the two extremes? Clarify the calculation and whether it is averaged over sites.","section":"Section 3.1"},{"comment":"The feature-selection procedure uses SelectKBest with F-regression, which is univariate and may not capture interactions that tree-based models use. This limitation is not discussed; a brief note would be helpful.","section":"Section 2.3"}],"recommendation":"major_revision","confidential_remarks":"The paper has a well-structured workflow and provides open code, which is a positive feature. However, the main scientific claim (s/p over d for partially-filled d-band TMs) hinges on a dataset-composition confound that the authors themselves acknowledge. The DFT setup also lacks basic validation for magnetic systems. These issues are reparable within the manuscript's scope, but they require either additional calculations (convergence tests, d-band-only baseline, TM-subset analysis) or substantial rewording of the claims. I recommend major revision rather than rejection, because the underlying data and ML pipeline are useful and the authors' acknowledgment of the confound suggests they can address it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely useful dataset and a fair-minded ML descriptor study, but the paper's central 'beyond d-band' claim is not yet established for the partially filled d-band alloys it emphasizes. It needs an ablation and convergence checks before that claim can be trusted.\n\nWhat's actually new: ~1200 adsorption energies across all surface sites of 47 IMCs, which is more systematic than prior d-band-only ML screening. The scaling-relation analysis showing NH3 decoupled from N2/N2H on multimetallic sites is a solid finding. The ML pipeline is reproducible (code on GitHub) and the SHAP analysis is interpreted honestly—the authors explicitly acknowledge that ~800/1200 configurations contain Al, Zn, Cu, Pt where s/p importance is expected. That honesty deserves credit.\n\nThe problem is that the headline conclusion goes beyond what the evidence shows. The global SHAP ranking cannot separate the s/p effect in partially filled d-band TMs from the effect of p-block/filled-d-band majority. The Fe-Co PDOS analysis (Fig. 9) is only two surfaces and qualitative; it shows sp-d hybridization but doesn't demonstrate that s/p features are more important than d-band for those systems. On top of that, no d-band-only baseline was trained, so 'more prominent' is not quantified as an improvement. The DFT settings are under-specified: 1x1x1 k-points for optimization, no spin-polarization statement for magnetic Fe/Co/Ni, and no vacuum/slab or thickness details. For adsorption energies that drive the catalyst ranking, this is a real gap; PBE with 1x1x1 can shift adsorption energies by tenths of eV, which could alter the Fe-Co recommendation. The energy windows for ranking are literature-defined, which is fine, but they inherit the DFT uncertainty.\n\nNone of this is fatal. The dataset and the empirical finding that s/p descriptors carry weight in the full dataset are real contributions. But the title and abstract overstate the generality. This is a solid computational screen with verification gaps, not a paradigm shift.\n\nWho is this for: people screening multimetallic NRR catalysts, and anyone working on descriptor-based ML for adsorption. It deserves a serious referee: the dataset is valuable and the questions are important, but the manuscript needs major revision (convergence tests, spin treatment, d-band-only ablation, and subset SHAP for TM-only systems) before the main claim is supported.\n\nRecommendation: send to peer review, with the expectation of substantial revision.","headline":"Useful new IMC adsorption dataset and an honest ML descriptor analysis, but the 'beyond d-band' claim is not yet proven for partially filled d-band alloys—needs ablation and convergence checks.","tokens_in":24559,"tokens_out":4350,"would_cite":true,"duration_ms":39223,"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 argues that s- and p-band electronic descriptors, not the traditional d-band center, control how strongly nitrogen-reduction intermediates bind on intermetallic surfaces, and that the Fe–Co compounds Fe9Co7 and Fe3Co are the best","keywords":["nitrogen reduction reaction","intermetallic compounds","d-band center","s- and p-band descriptors","density functional theory","machine learning","scaling relationships","ammonia synthesis"],"falsifier":"Recompute the adsorption energies for Fe3Co and Fe9Co7 with spin-polarized DFT, a denser k-point mesh, and converged slab thickness; if the site ordering or the N2/N2H/NH3 adsorption windows shift by more than a few tenths of an eV, or if the predicted sp–d hybridization peaks in the projected density of states disappear, the central descriptor claim and the catalyst ranking would fail.","tokens_in":23610,"feed_emoji":"🧪","tokens_out":6647,"duration_ms":62315,"temperature":0.7,"pith_summary":"This study screens 47 bimetallic intermetallic compounds for the nitrogen reduction reaction, computing adsorption energies of N2, N2H, and NH3 at roughly 1,200 surface sites. It claims that the conventional d-band center is not the dominant descriptor on these multimetallic surfaces; s- and p-band centers and fillings, along with the size and electronegativity of atoms neighboring the adsorption site, carry more predictive weight, especially for N2 and N2H. The authors argue this reflects adsorption-induced sp–d hybridization that operates even in transition metals with partially filled d-bands, not just in p-block or nearly filled d-band systems. If right, the compact 20-feature descriptor set and the Fe–Co candidates Fe9Co7 and Fe3Co give a concrete, noble-metal-free path toward ammonia synthesis catalyst design.","feed_headline":"s and p bands, not d band, set nitrogen-reduction strength","feed_subtitle":"A DFT–ML screen of 47 intermetallics points to Fe–Co alloys that balance N2 activation with NH3 release.","key_machinery":"The carrying mechanism is a descriptor-based machine-learning pipeline built from per-site DFT data: for each adsorption configuration, 72 features encode s-, p-, and d-band centers and fillings (local and global), atomic charges of neighboring atoms, intrinsic atomic properties of the binding atom (electronegativity, electron affinity, covalent radius, ionization energy, and related quantities), and six aggregated statistics over shells of up to 20 neighboring atoms. Feature selection reduces the set to 20; gradient-boosted tree models then predict adsorption energies with mean absolute errors of 0.26 eV for N2, 0.39 eV for N2H, and 0.17 eV for NH3. A feature-attribution analysis ranks desc","core_discovery":"The central claim: on intermetallic surfaces, adsorption strength for nitrogen-reduction intermediates is controlled more by s- and p-band features and the local atomic environment than by the d-band center. Using DFT energies for about 1,200 adsorption configurations and ML models reduced to 20 features, the authors find that local s- and p-band centers, the maximum covalent radius in the neighbor shell, and shell electronegativity dominate predictions for N2 and N2H, while the d-band center is not among the top features. Mechanistically, they argue that adsorption induces sp–d hybridization—co-localized, simultaneous modification of sp and d states in the projected density of states—even i","pith_inferences":["Editorial inference: if the sp–d hybridization picture generalizes, descriptor-based screening for other reactions on intermetallics—such as CO2 electroreduction, hydrogen evolution, or oxygen reduction—should include s- and p-band features rather than assuming the d-band center is sufficient.","Editorial inference: because most of the roughly 1,200 configurations contain Al, Zn, Cu, or Pt, elements where s- and p-band roles are already expected, the strongest independent evidence for the partially-filled-d-band claim is the projected-DOS analysis of just Fe3Co and Fe9Co7; extending the claim broadly would require testing more transition-metal-only intermetallics.","Editorial inference: the broken NH3–N2H scaling relationship suggests a testable design rule—increase the asymmetry of nearest-neighbor electronegativity at active sites to decouple activation from product release.","Editorial inference: an experimental test on synthesized Fe9Co7 or Fe3Co nanoparticles, measuring N2 adsorption isotherms or ammonia synthesis rates, would directly probe whether the computed adsorption windows translate into catalytic activity."],"forward_implications":["Fe9Co7 and Fe3Co emerge as promising non-noble NRR catalysts because they fall in the adsorption-energy windows for N2, N2H, and NH3 simultaneously.","The d-band center alone is insufficient for screening intermetallic catalysts; s- and p-band features and local environment descriptors should be included in descriptor sets.","The decoupling of NH3 from the N2–N2H scaling relationship on multimetallic sites opens the possibility of independently optimizing nitrogen activation and ammonia desorption.","A compact 20-feature set with simple, computationally efficient ML models can predict adsorption energies accurately enough for rapid screening of additional intermetallic compounds.","The role of s- and p-orbitals extends beyond p-block elements and nearly filled d-band metals to transition metals with partially filled d-bands, via adsorption-induced sp–d hybridization."],"fun_headline_variants":["D-band dethroned: s and p orbitals control NRR adsorption","ML screen: s-p bands beat d-band for N2 reduction catalysts","Fe-Co alloys emerge via s/p-band descriptors, not d-band","NRR adsorption: s,p bands trump d-band in intermetallics"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the DFT adsorption energies—computed with a single k-point for geometry optimization, a generalized-gradient functional, and no reported spin-polarized treatment or slab-thickness convergence—are accurate enough to rank surface sites on magnetic Fe, Co, and Ni surfaces, where errors of a few tenths of an eV could change both the catalyst ranking and the feature-importance conclusions.","fun_headline_variants_meta":{"raw":{"variants":["D-band dethroned: s and p orbitals control NRR adsorption","ML screen: s-p bands beat d-band for N2 reduction catalysts","Fe-Co alloys emerge via s/p-band descriptors, not d-band","NRR adsorption: s,p bands trump d-band in intermetallics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000675,"raw_usage":{"total_tokens":2959,"prompt_tokens":846,"completion_tokens":2113,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":2033}},"tokens_in":590,"tokens_out":2113,"duration_ms":13326,"temperature":1.0,"reasoning_tokens":2033,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T17:51:40.033173+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the adsorption energies for Fe3Co and Fe9Co7 with spin-polarized DFT, a denser k-point mesh, and converged slab thickness; if the site ordering or the N2/N2H/NH3 adsorption windows shift by more than a few tenths of an eV, or if the predicted sp–d hybridization peaks in the projected density of states disappear, the central descriptor claim and the catalyst ranking would fail.","supporting_citations":[],"review_version":1}