{"id":"56018ee1-4caf-4d02-a3de-d2e0b58b310c","arxiv_id":"2412.13838","paper_version":5,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A new descriptor, the adsorption energy distribution, clusters 158 alloys by similarity to known CO2-to-methanol catalysts and proposes ZnRh and ZnPt3 as promising untested candidates.","lead":"This paper introduces a new way to describe metal catalysts, by turning the spread of adsorption energies across many surfaces and binding sites into a fingerprint called the adsorption energy distribution. It uses machine-learned force fields to compute these fingerprints for 158 alloys and clusters them to suggest two untested alloys, ZnRh and ZnPt3, as candidates for converting CO2 into methanol.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The candidate nomination hinges on cluster membership that is never checked against the MLFF's own stated adsorption-energy error; with a Wasserstein merge threshold of 2.5e-3 eV and validation MAEs of 0.06-0.16 eV, a modest energy error could move ZnRh or ZnPt3 out of cluster 10.","rationale":"The paper's central deliverable is a screening prediction, and the screening step is the clustering. The authors are transparent about the AED-activity assumption, so the most damaging correctable weakness is whether the cluster assignment survives the errors the paper itself reports. The paper validates the MLFF on only three materials and uses a 0.25 eV EMAE cutoff for inclusion; no uncertainty analysis is applied to the final clustering. Since W1 distance has units of energy, the merge threshold of 0.0025 eV is far smaller than the reported adsorption-energy error bars, making the dendrogram potentially unstable. This is a quantitative, internal risk rather than a disagreement with the field's consensus. The proposed bootstrap test would resolve it. If it passes, the paper's conditional status remains justified; if it fails, the candidate recommendations should be re-evaluated. The reader's weakest_assumption is related but different: it concerns semantics of similarity, while ours concerns accuracy of the similarity calculation; hence partial agreement. The workflow is otherwise reproducible, with open data, a validation protocol, and honest limitations, so the verdict should remain CONDITIONAL rather than being tightened or loosened.","tokens_in":17832,"tokens_out":5222,"duration_ms":50410,"concrete_test":"Bootstrap perturbation test: for every retained material, draw 1000 realizations of its four AEDs by adding per-adsorbate Gaussian noise with sigma equal to that adsorbate's EMAE (or 0.16 eV where unavailable), re-bin with the same 0.1 eV bins, recompute the Wasserstein matrix and Ward clustering at threshold 2.5e-3, and measure the co-membership frequency of ZnRh and ZnPt3 with CuZn, NiZn, and Ga2Cu. If the frequency is below 95%, the central nomination is not robust to stated MLFF error; if above 95%, this specific concern is settled and the remaining risk is the acknowledged AED-activity assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.5 sets the agglomerative clustering threshold at 2.5e-3 eV on Wasserstein distances computed from 0.1 eV-binned AEDs, while Section 2 reports validation MAEs of 0.06 eV (Pt), 0.12 eV (Zn), 0.06 eV (NiZn) and an overall 0.16 eV for the MLFF. W1 distance is translation-sensitive: a uniform +0.1 eV error in one material's adsorption energies changes its W1 to an unshifted material by about 0.1 eV, two orders of magnitude above the merge threshold. The paper excludes materials with EMAE > 0.25 eV but does not report EMAEs for ZnRh/ZnPt3, nor does it propagate validation errors into the distance matrix or cluster assignments. Therefore the claim that ZnRh and ZnPt3 cluster with CuZn/NiZn/Ga2Cu is not grounded in the stated accuracy of the descriptor inputs; it may be an artifact of treating MLFF predictions as exact. The problem is internal to the workflow, not the conceptual AED-activity assumption, and is checkable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a new catalytic descriptor, the adsorption energy distribution (AED), which aggregates machine-learned force field (MLFF) adsorption energies for four reaction intermediates across many surface facets and binding sites of ~160 metallic alloys. The authors validate the OCP equiformer_V2 MLFF on Pt, Zn, and NiZn, then use a sparse single-point DFT validation protocol (EMAE) to estimate errors and exclude 29 materials with EMAE above 0.25 eV. They compute Wasserstein distances between the AEDs and perform hierarchical clustering with Ward linkage, identifying ZnRh and ZnPt3 as untested candidate catalysts because they cluster with known active materials such as CuZn, NiZn, and Ga2Cu. The paper is transparent about the central assumption that AED similarity to a known good catalyst is a meaningful activity indicator, and it makes the resulting dataset available on Zenodo.","tokens_in":18138,"tokens_out":4558,"duration_ms":42159,"significance":"If the workflow is robust, the AED descriptor is a genuinely new way to represent facet and site heterogeneity in catalyst screening, and it is reaction-tunable through adsorbate choice. The dataset of over 877,000 adsorption energies and the open data publication are concrete strengths, as is the explicit EMAE validation protocol. The candidate nominations, however, rest on cluster assignments that are never checked against the stated MLFF accuracy, and on an untested similarity-to-activity assumption; the former is an internal consistency issue that is checkable and fixable, while the latter limits the strength of the claims that can be drawn.","major_comments":[{"comment":"The clustering threshold of 2.5×10−3 eV is not reconciled with the MLFF validation MAEs of 0.06–0.16 eV reported in Table 1 or with the 0.25 eV EMAE cutoff used for data cleaning. For the first Wasserstein distance in 1D, a uniform shift of one distribution by δ changes its distance to an unshifted distribution by approximately δ. A systematic MLFF error of 0.1 eV in the adsorption energies of one material therefore shifts its W1 distance to another material by roughly 0.1 eV, which is 40 times larger than the merge threshold. This means that the placement of ZnRh and ZnPt3 in cluster 10, and hence the central candidate claim, may be an artifact of treating MLFF energies as exact. The authors should report EMAEs for ZnRh and ZnPt3, propagate MLFF uncertainties into the distance matrix (for example by bootstrapping adsorption energies with the measured MAE), and demonstrate that the cluster assignments are stable under perturbations of the order of the validation MAE.","section":"§3.5 / §2 (Validation and Data Cleaning)"},{"comment":"The paper states explicitly that the candidate proposals rest on the assumption that AED similarity to a known good catalyst indicates promising catalytic activity, but this assumption is not validated by any independent test. To support the specific claim that ZnRh and ZnPt3 'should have a high activity,' the authors should either check whether experimental activity rankings among the known catalysts in cluster 10 are reproduced by AED distances, or perform microkinetic modeling for the proposed candidates, or clearly label the candidates as untested hypotheses. As written, the recommendation is stronger than the evidence supports.","section":"§2 (Unsupervised Learning)"},{"comment":"The exclusion of 29 materials with EMAE above 0.25 eV is a post hoc cleaning step whose relation to the clustering is unclear. The 0.25 eV threshold is two orders of magnitude larger than the clustering threshold of 2.5×10−3 eV, yet the EMAEs of the retained materials, at least for ZnRh and ZnPt3, are not reported in the text or in the supplement; the reader cannot verify that the AEDs of the proposed candidates are within the accuracy needed to support their cluster assignment. The paper should justify the 0.25 eV cutoff in terms of its effect on the downstream Wasserstein distances and report the relevant EMAE values for the proposed candidates.","section":"§2 (Validation and Data Cleaning)"}],"minor_comments":[{"comment":"The caption uses 'EMEA' where 'EMAE' is meant.","section":"Fig. 2 caption"},{"comment":"References [64] and [65] are identical (both cite Ong et al., pymatgen); one should be replaced with the correct Custodian reference.","section":"References"},{"comment":"The phrase 'the price energy zero is of no relevance' appears to be a typo for 'the precise energy zero is of no relevance'.","section":"§2 (Adsorption Energy Distributions)"},{"comment":"The caption contains 'absorbate-material combination', which should be 'adsorbate-material combination'.","section":"Table 1 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is transparent about its limitations and the data availability is a clear strength. The main issue is that the central candidate prediction rests on a clustering analysis whose sensitivity to the stated MLFF errors is not characterized; this is an internal consistency problem that is checkable and fixable within a revision. The authors should be encouraged to add an error-propagation analysis and to temper the candidate claims accordingly. The duplicate references 64/65 suggest a careful proofread is also needed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is worth your time if you care about high-throughput catalyst screening. The genuinely new piece is the adsorption energy distribution (AED) descriptor: facet- and site-resolved adsorption energies are collected into normalized histograms, concatenated across four CO2-to-methanol intermediates, and compared between materials via Wasserstein distance. That extends Sabatier-type reasoning beyond single facets in a sensible way. The workflow is also solid: roughly 877k adsorption energies for 158 alloys from the OCP MLFF, DFT single-point validation, an EMAE-based rejection step, and the full dataset on Zenodo. That reproducibility is real credit.\n\nThe authors state their key limitation plainly: the candidate predictions for ZnRh and ZnPt3 rely on the assumption that AED similarity to known active catalysts means catalytic activity. They have not validated that with microkinetics or experiments, and they do not pretend otherwise.\n\nThe soft spot that matters more is internal and checkable. The hierarchical clustering uses a Wasserstein merge threshold of 2.5e-3 eV, while the MLFF validation MAE on Pt, Zn, and NiZn is 0.06-0.16 eV. W1 distance moves linearly with a uniform shift in adsorption energies, so a 0.1 eV MLFF error can move a material's distance by ~0.1 eV—orders of magnitude above the merge threshold. The paper excludes materials with EMAE above 0.25 eV but never reports the EMAEs for ZnRh and ZnPt3, and never propagates any error estimate into the distance matrix or cluster assignments. So the headline claim that these two alloys cluster with CuZn and Ga2Cu may be an artifact of treating the MLFF as exact. This is fixable—a sensitivity analysis or a bootstrap over the error distribution would settle it—but it is load-bearing.\n\nSmaller issue: the 0.25 eV EMAE cutoff is applied post hoc and excludes 29 mostly magnetic materials. That is defensible, but it is a systematic narrowing of the search space, not merely noise removal.\n\nBottom line: honest, useful, and reproducible, with a new descriptor that deserves attention. The candidate predictions need the clustering sensitivity check before I would trust them. I would send this to a serious referee, and I would cite the AED method even while treating the specific candidates as provisional.","headline":"A sensible, honestly limited screening study whose two candidate predictions rest on a clustering step that the paper never checks against its own MLFF error bars.","tokens_in":18593,"tokens_out":2560,"would_cite":true,"duration_ms":22202,"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":"The paper proposes a new catalyst descriptor—the distribution of adsorption energies across facets and sites—and uses it to flag ZnRh and ZnPt3 as promising, untested candidates for CO2-to-methanol conversion.","keywords":["CO2 hydrogenation","methanol synthesis","adsorption energy distribution","machine-learned force fields","catalyst discovery","hierarchical clustering","Wasserstein distance","bimetallic alloys"],"falsifier":"Measure the methanol yield, selectivity, and stability of ZnRh and ZnPt3 under standard CO2 hydrogenation conditions (roughly 800 K, H2/CO2 feed) and compare with CuZn or industrial Cu/ZnO/Al2O3; if their performance is not in the same league despite AED similarity to known active catalysts, the descriptor's central assumption fails.","tokens_in":17682,"feed_emoji":"🧪","tokens_out":8557,"duration_ms":74245,"temperature":0.7,"pith_summary":"This paper tries to establish a fast, general route to catalyst discovery: characterize a material not by one binding energy but by the whole distribution of adsorption energies of key reaction intermediates across all its surface facets and binding sites. Using a pre-trained machine-learned force field, the authors compute these adsorption-energy distributions for nearly 160 metallic and bimetallic alloys relevant to CO2-to-methanol conversion, generating over 877,000 adsorption energies. They then compare the distributions with the Wasserstein distance, cluster the materials hierarchically, and find that many known active catalysts—Cu, CuZn, NiZn, Ga2Cu, InPt3, Ni—fall into one cluster. On that basis they propose ZnRh and ZnPt3, which sit in the same cluster but have not been tested, as likely high-activity candidates, with melting points above Cu/CuZn that suggest better stability under reaction conditions.","feed_headline":"Untested alloys ZnRh and ZnPt3 flagged as CO2-to-methanol catalysts","feed_subtitle":"Their adsorption-energy fingerprints match CuZn, Ga2Cu, and NiZn—and their melting points are higher.","key_machinery":"The load-bearing object is the adsorption energy distribution (AED): for each material, a normalized histogram with 0.1 eV bins covering the adsorption energies of four intermediates (*H, *OH, *OCHO, *OCH3) across all generated low-Miller-index surface terminations and adsorption sites, with padding so the four adsorbate distributions do not overlap. The AED replaces the usual practice of quoting a single surface minimum or a d-band center. Comparison is done with the first Wasserstein distance, which measures how much probability mass must be moved to turn one distribution into another, followed by agglomerative hierarchical clustering with variance-minimizing linkage. A companion statistical summary (SAAED) condenses each AED into minimum, spread, and higher moments so individual binding energies can be checked against the Sabatier principle, for example against the oxygen-binding volcano for Cu-based methanol synthesis.","core_discovery":"The paper's central claim is that an adsorption energy distribution (AED)—a normalized histogram of adsorption energies for *H, *OH, *OCHO, and *OCH3 collected over multiple Miller-index surfaces and symmetry-distinct sites—can serve as a material fingerprint for catalytic activity in CO2-to-methanol conversion. After validating the machine-learned force field against density functional theory on Pt, Zn, and NiZn and excluding materials whose estimated error exceeds 0.25 eV, the authors build AEDs for 158 materials. Hierarchical clustering of the Wasserstein distance matrix yields 19 clusters, one of which (cluster 10) groups known active catalysts with ZnRh and ZnPt3. The paper therefore proposes ZnRh and ZnPt3 as promising candidates that should have high activity and, because their melting temperatures exceed those of Cu and CuZn, possibly better durability.","pith_inferences":["A natural next step is to test whether cluster membership is robust to the choice of adsorbates; adding CO or other intermediates could shift ZnRh and ZnPt3 relative to the known catalysts and would reveal which energy matches are doing the work.","Because the descriptor averages over facets without weighting them by their surface areas, combining AEDs with Wulff-construction facet weights could sharpen the comparison for nanoparticles whose exposed facets change under reaction conditions.","If the similarity assumption holds, the same clustering logic could be applied across reaction families: a catalyst found active for CO2-to-methanol might seed searches for methanation or higher-alcohol synthesis by using the appropriate intermediate set."],"forward_implications":["ZnRh and ZnPt3 become concrete experimental targets: if they show methanol activity comparable to CuZn or NiZn, the AED-similarity route is a validated screening tool.","The same pipeline can be pointed at other reactions by swapping the four adsorbates for that reaction's key intermediates, so the descriptor is not limited to CO2-to-methanol.","Materials with broad AEDs, such as Y3Zn11, V6Ga5, Y3In5, and V4Zn5, are predicted to be poorer catalysts because only a small fraction of their surface binds intermediates near the optimum.","The melting-point argument implies that ZnRh and ZnPt3 may retain stability where Cu-based catalysts degrade, which is directly testable in long-run reactor tests.","Excluding 29 materials with large estimated errors—many of them magnetic—means the current search space is biased against magnetic alloys; including spin polarization could extend the method."],"supporting_citations":[{"why":"Identifies the four surface-bound intermediates and the Cu(211) reference surface used to build the AEDs.","marker":"[30]"},{"why":"Supplies the training data and adsorption-energy convention behind the pre-trained machine-learned force field.","marker":"[33]"},{"why":"Defines the Wasserstein distance used to compare adsorption-energy distributions in the clustering.","marker":"[35]"},{"why":"Delimits the initial element search space from experimental CO2-reduction studies.","marker":"[36]"},{"why":"Supplies the stable experimentally observed bulk structures used as the starting point for surface generation.","marker":"[37]"},{"why":"Reports the claimed accuracy of the machine-learned force field that the validation protocol is benchmarked against.","marker":"[42]"},{"why":"Introduces the variance-minimizing hierarchical clustering criterion used to define the material clusters.","marker":"[44]"},{"why":"Reports NiZn as an effective CO2-to-methanol catalyst, one of the known active materials grouped in cluster 10.","marker":"[45]"},{"why":"Provides the oxygen-binding volcano plot connecting *OH, *OCHO, and *OCH3 adsorption energies to methanol activity on Cu-based surfaces.","marker":"[49]"}],"fun_headline_variants":["Machine-learned fingerprint flags ZnRh and ZnPt3 for methanol synthesis","New descriptor predicts ZnRh, ZnPt3 as promising CO2-to-methanol catalysts","ML screens 158 alloys, highlights ZnRh and ZnPt3 for methanol","Adsorption-energy descriptor finds stable untested alloys for methanol","ZnRh and ZnPt3: untested alloys with catalyst potential from ML screen"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The assumption that carries the load is that two materials with similar adsorption-energy distributions will have similar catalytic activity; the paper states this explicitly and does not validate it with experiment or microkinetic modeling.","fun_headline_variants_meta":{"raw":{"variants":["Machine-learned fingerprint flags ZnRh and ZnPt3 for methanol synthesis","New descriptor predicts ZnRh, ZnPt3 as promising CO2-to-methanol catalysts","ML screens 158 alloys, highlights ZnRh and ZnPt3 for methanol","Adsorption-energy descriptor finds stable untested alloys for methanol","ZnRh and ZnPt3: untested alloys with catalyst potential from ML screen"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000665,"raw_usage":{"total_tokens":3005,"prompt_tokens":887,"completion_tokens":2118,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":2019}},"tokens_in":503,"tokens_out":2118,"duration_ms":14307,"temperature":1.0,"reasoning_tokens":2019,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:43:36.364195+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the methanol yield, selectivity, and stability of ZnRh and ZnPt3 under standard CO2 hydrogenation conditions (roughly 800 K, H2/CO2 feed) and compare with CuZn or industrial Cu/ZnO/Al2O3; if their performance is not in the same league despite AED similarity to known active catalysts, the descriptor's central assumption fails.","supporting_citations":[{"cited_title":"Journal of Cleaner Production 339, 130653 (2022) https://doi.org/10.1016/j.jclepro.2022.130653","cited_arxiv_id":null,"evidence_quote":"Delimits the initial element search space from experimental CO2-reduction studies."},{"cited_title":"Nature Chemistry 6(4), 320 (2014) https://doi","cited_arxiv_id":null,"evidence_quote":"Provides the oxygen-binding volcano plot connecting *OH, *OCHO, and *OCH3 adsorption energies to methanol activity on Cu-based surfaces."}],"review_version":1}