{"id":"e695348a-dbfe-4ffd-a600-6e5abdb84e4d","arxiv_id":"2502.04603","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A database and random-forest models for 1,147 Janus 2D material-metal heterostructures show that bulk substrate electronegativity and surface energy mostly control binding and spacing.","lead":"This paper reports high-throughput simulations of nearly 1,000 Janus 2D material on bulk metal heterostructures and machine-learning models that predict how strongly they bind. The models and open database could help researchers pick substrates that stabilize promising 2D materials for electronics and sensing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Bulk-dominance claim is preconditioned by sampling: 2D electronegativity spans only 1.98-2.59 versus bulk 0.95-2.54, and all 19 substrates are metallic (111); feature importance may reflect range mismatch, not a fundamental factor.","rationale":"The reader's weakest-assumption identification matches my own: the bulk-dominance conclusion is inferred from a dataset where the 2D side has artificially low variance in the key descriptor (electronegativity) and where only 19 metallic cubic (111) substrates are sampled. This is the single most load-bearing concern because the paper's strongest claim is explicitly a statement about which material side fundamentally controls interface stability. If the feature-importance ranking changes under a variance-matched or broader test, the abstract's general phrasing would overstate the finding, even though the DFT data and ML models for the studied subspace are credible. I considered other potential concerns, such as the use of impurity-based feature importance, the stability language around ΔEf_ads, and possible cross-validation leakage, but none is as directly tied to the central claim as the sampling precondition. The authors' own in-text caveat supports this reading, so the concern is not manufactured. The reader's CONDITIONAL verdict is appropriate: the paper makes a real contribution (new database, 438 relaxed interfaces, reproducible pipeline) but should not claim a universal 'fundamental factor' until the sampling bias is tested. Thus I recommend no change to the reader's verdict.","tokens_in":13408,"tokens_out":5962,"duration_ms":67009,"concrete_test":"Using the published 438-pair dataset (aiHD), construct a variance-matched subsample in which the bulk electronegativity range is restricted to the 2D range (1.98-2.59) by excluding the most electropositive substrates (e.g., Sr, Y, Sc, Hf) or by stratified undersampling; retrain the exact RFR_bind and RFR_z-sep pipelines and recompute permutation feature importance. If bulk electronegativity (or surface energy) ceases to be the top feature, or if 2D electronegativity or electronegativity-difference importance rises above it, the reported bulk dominance is an artifact of imbalanced feature ranges rather than a fundamental factor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (abstract and 'Machine Learning Insights' section) is that bulk electronegativity and bulk surface energy are the leading fundamental factors controlling binding energy and z-separation. For this to be a general statement, the random-forest feature-importance ranking must reflect physical causation rather than sampling leverage. The dataset is preconditioned in two ways: (i) the 51 Janus 2D materials are all Y-M-X compounds with Y=S/Se/Te, M=As/Bi/Cr/Hf/Mo/Nb/Sb, X=Br/Cl/I/Se/Te, so their composite electronegativities fall in the narrow window 1.98-2.59, while the 19 bulk metals span 0.95-2.54; (ii) substrates are limited to metallic cubic elements on (111) surfaces selected by lattice mismatch <3% and cell area <80 Å^2. Under these conditions, bulk electronegativity has far more variance to explain ΔEb, and bulk surface energy similarly dominates z-separation. The authors explicitly acknowledge the narrow 2D range ('Thus, for this subset of materials, the bulk electronegativity plays a much larger role than the difference') and even recommend studying 2D materials with a larger electronegativity range, yet the abstract and conclusion state the dominance as a general finding without that caveat. The claim is therefore externally valid only for this chemical subspace; for 2D materials with higher or lower electronegativity (e.g., fluorides, oxides, alkali-based 2D compounds), the balance could shift. This is a correctness risk for the headline claim, but not an internal computational error.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a high-throughput DFT study of 1147 Janus 2D-bulk heterostructures formed from 51 Janus 2D materials and 19 metallic cubic (111) substrates, together with random-forest regression models that predict the binding energy and interfacial z-separation of the most stable configuration for each of 438 2D-bulk pairs. The reported models achieve RMSEs of 0.047 eV/atom and 0.139 Å, respectively. The central physical claim is that bulk-material properties, specifically bulk electronegativity for binding energy and bulk surface energy for z-separation, are the leading fundamental factors governing heterostructure stabilization, implying that substrate choice is the primary design lever. The computed data are released in the aiHD database.","tokens_in":13743,"tokens_out":4781,"duration_ms":45092,"significance":"If the central claim holds, the work is significant: it provides a large, openly available database of ab initio 2D-bulk interface data, demonstrates that random-forest models can capture trends in binding energy and interfacial separation across this chemical space, and offers a physically plausible case study of charge redistribution and hybridization at the SbTeI/Sr interface. Strengths include the consistent vdW-corrected DFT workflow, the large dataset relative to prior 2D-bulk studies, the explicit cross-validation of the ML models, and the authors' own acknowledgment of the narrow 2D electronegativity range in the feature-importance discussion. However, the generalization of the feature-importance conclusion is not yet fully supported because the dataset is deliberately narrow on the 2D side and the feature-importance rankings are sensitive to that sampling.","major_comments":[{"comment":"The headline claim that bulk electronegativity and bulk surface energy are the 'leading fundamental factor' is not established by the reported feature-importance analysis. Random-forest feature importance depends on the range and variance of features in the training set. Here the Janus 2D electronegativities lie between 1.98 and 2.59, while the bulk values span 0.95 to 2.54, and all 19 substrates are metallic cubic (111) surfaces. Under this sampling, the bulk electronegativity has far more variance to explain ΔEb, so its high importance is partly a sampling artifact. The authors acknowledge this limitation for electronegativity ('Thus, for this subset of materials, the bulk electronegativity plays a much larger role than the difference'), yet the abstract and conclusion state the dominance as a general finding without that caveat. A concrete control is needed: for example, retrain the models on a set with a broader 2D electronegativity range (e.g., oxides, fluorides, or alkali-based 2D compounds) or use permutation importance on variance-balanced resamples. If bulk dominance persists under such controls, the claim is strengthened; otherwise, it should be restricted to the studied chemical subspace.","section":"Machine Learning Insights into the Fundamental Factors Governing Janus 2D Heterostructure Stability and Figure 7(b)"},{"comment":"The term 'thermodynamically stable' overstates what is computed. The criterion ΔEf_ads < 0 compares the heterostructure with the freestanding 2D material and its 3D bulk counterpart; it does not establish stability against all competing phases and includes no temperature or pressure dependence. The 828 'stable' configurations are better described as 'bulk-stabilized relative to the freestanding 2D material.' Because the number 828 is cited in the abstract as a central result, the wording should be corrected to match the actual definition.","section":"Table 1 and Section 'Energetic Stability of Janus 2D Materials'"}],"minor_comments":[{"comment":"The reported RMSE and MAE values are cross-validated on the same dataset, but no uncertainty intervals are given. Reporting standard deviations across cross-validation folds would help assess whether the differences between model variants (e.g., RFRz−sep and RFRz−sep*) are meaningful.","section":"Table 1 and Figure 7"},{"comment":"The feature-reduction process is described only at a high level, and the final lists of 15 and 10 features are not given in the main text. Since the feature set is central to the interpretation, the full feature lists and their importances should be included or more completely summarized in the main text.","section":"Machine Learning Methodology"},{"comment":"The abstract says 'nearly 1000 heterostructures' but the paper reports 1147 heterostructure configurations; please make the count consistent.","section":"Abstract"},{"comment":"There is a typo in 'V ASP' (should be VASP) and in 'M atminer's database' (should be Matminer) in the text near Figure 3.","section":"DFT Methods"},{"comment":"The phrase 'the lower the ΔEb the lower the stability' is confusing: since ΔEb is defined as (E2D + ES − E2D+S)/N2D, a lower (more negative) ΔEb means stronger binding and thus higher stability. Please clarify the sign convention.","section":"Energetic Stability of Janus 2D Materials"}],"recommendation":"major_revision","confidential_remarks":"This is a useful dataset contribution with a sound DFT workflow. The main issue is that the feature-importance conclusion is overgeneralized relative to the sampling; this is fixable by tempering the claim or adding a variance-balanced control. I would not recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper is worth engaging. The new asset is aiHD, the first open DFT database of Janus 2D-bulk heterostructures, with nearly a thousand relaxed interfaces and an ML model trained on them. The DFT workflow is standard but solid, and the random-forest metrics (RMSE 0.047 eV/atom for binding energy, 0.139 Å for z-separation) are decent for such a small dataset.\n\nThe feature-importance analysis is the part to read with a grain of salt. The authors claim bulk electronegativity and bulk surface energy dominate, and the abstract and conclusion state that as a general fact. But the dataset is deliberately narrow: 51 Janus 2D materials with electronegativities squeezed between 1.98 and 2.59, versus 19 metallic cubic (111) substrates spanning 0.95 to 2.54. With that imbalance, it is almost guaranteed that bulk properties will carry most of the variance. The authors actually acknowledge this in the ML section—'for this subset of materials, the bulk electronegativity plays a much larger role'—but then the abstract drops the qualifier. That is an overgeneralization, not a computational error. A motivated referee could test it by including, say, oxide or halide 2D materials with higher electronegativity.\n\nOther soft spots are minor. The 'stabilization' language refers only to the adsorption energy relative to the isolated 2D material, not to thermodynamic stability of the heterostructure as a phase. And the RMSE values come from K-fold CV, so they are honest in-sample estimates, but no error bars or holdout test is reported. Those are fixable in revision.\n\nThe paper deserves a serious referee. The database itself is a real contribution and the ML model is useful for substrate screening in this chemical space. I would recommend acceptance after the authors temper the generality claims and add a short paragraph on the sampling limits. If they refuse to soften the abstract, that would be a reason to push back.","headline":"A genuinely useful dataset and model, but the bulk-dominance conclusion is oversold beyond the sampled chemical space.","tokens_in":14278,"tokens_out":1805,"would_cite":true,"duration_ms":18616,"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":"Bulk properties, not the 2D layer, set heterostructure stability.","keywords":["Janus 2D materials","heterostructure stability","random forest regression","binding energy","z-separation distance","bulk surface energy","electronegativity","high-throughput DFT"],"falsifier":"Extend the study to non-Janus 2D materials with electronegativities below 1.5 or above 3.0 on the same 19 substrates, recompute binding energies, and retrain the random forest; if 2D electronegativity or the 2D–bulk difference enters the top features, the bulk-dominance claim fails for the broader class.","tokens_in":13228,"feed_emoji":"🔬","tokens_out":7161,"duration_ms":62855,"temperature":0.7,"pith_summary":"The paper asks which partner controls the stability of a three-atom-thick Janus 2D material placed on a metal surface. Based on 1147 density-functional-theory heterostructures and random-forest models, the answer is the bulk support: bulk electronegativity is the leading predictor of binding energy, and bulk surface energy is the leading predictor of the interfacial z-separation. The models reproduce the DFT data with root-mean-square errors of 0.047 eV/atom and 0.139 Å, and the calculated trends align with experimental observations of which metal substrates grow transition-metal dichalcogenides successfully. If this holds, substrate selection becomes the primary design lever for stabilizing and tuning 2D-on-bulk interfaces.","feed_headline":"Bulk properties, not the 2D layer, set heterostructure stability","feed_subtitle":"Random-forest models predict binding energy to 0.047 eV/atom using the support's properties alone.","key_machinery":"The machinery is a pair of random-forest regression models trained on the lowest-energy DFT configuration for each 2D-bulk pair, using 155 candidate features from the 2D material, the bulk slab, and the relaxed interface. Feature reduction (low-variance, high-correlation, recursive elimination) leaves 15 features for the binding-energy model and 10 for the z-separation model. A deliberately restricted z-separation model uses only features available from databases and elemental properties—bulk surface energy, atomic number, packing efficiency, molar volume—and still predicts separation with R2 of 0.838. The physical argument that bulk properties should dominate is that the bulk surface is created by breaking strong metallic bonds while the 2D material's exfoliation breaks weak van der Waals bonds, so the bulk side sets the interaction energy scale.","core_discovery":"The central claim is that in Janus 2D-bulk heterostructures, the bulk material's intrinsic properties dominate the interface energetics and geometry. For binding energy, bulk electronegativity has the highest feature importance (29%), ahead of the 2D–bulk electronegativity difference (13%), and removing the difference from the model slightly improves accuracy. For z-separation, bulk surface energy is the top feature, and higher surface energy correlates with smaller separation, consistent with more dangling bonds at the surface. The random-forest models, trained on the most stable configuration of each of 438 2D-bulk pairs, achieve R2 of 0.940 for binding energy and 0.838 for z-separation when the z-model is restricted to features that require no DFT calculation. The authors interpret the bulk dominance physically: creating the bulk surface breaks strong metallic bonds, whereas exfoliating the 2D layer breaks weak van der Waals bonds, so the bulk side carries the energy scale.","pith_inferences":["The bulk-dominated feature ranking may partly reflect the narrow electronegativity range of the 2D set (1.98–2.59) compared with the bulk set (0.95–2.54); including 2D materials with more extreme electronegativities could shift importance toward the 2D side.","The 19 elemental cubic (111) substrates are all metals, so the conclusion that bulk properties dominate may not transfer to polar, compound, or differently oriented surfaces without retesting.","Stability here is thermodynamic (negative adsorption formation energy); kinetic barriers to nucleation and growth are not addressed, so a predicted stable pair might still fail in synthesis.","The no-DFT z-separation model's modest accuracy drop (R2 0.886 to 0.838) suggests it is best used for coarse screening, with DFT reserved for finalists."],"forward_implications":["Substrate choice, not the 2D layer, becomes the primary design knob for stabilizing metastable Janus monolayers on metals.","The random-forest models can screen 2D-bulk pairs at negligible cost, with the z-separation model needing only database-available properties.","The published database of over 1200 heterostructures gives other researchers a benchmark and training set for interface models.","The computed stability ordering rationalizes experimental reports that some metal substrates (for example, Cu) grow many TMDCs while others (for example, Ni) grow few.","Because binding energy and z-separation correlate, a single bulk descriptor such as surface energy can guide targeted tuning of interface geometry."],"supporting_citations":[{"why":"Supplies the 51 Janus 2D structures and their computed properties from a public 2D materials database; this is the source of the 2D side of the data set.","marker":"[18]"},{"why":"Provides the high-throughput heterostructure construction framework and the DFT parameters and settings used to build and relax every 2D-bulk pair.","marker":"[27]"},{"why":"Supplies the bulk elemental structures and stability data along with analysis tools used to prepare the substrate set.","marker":"[31]"},{"why":"Provides the random-forest regression implementation and the cross-validation and feature-selection routines used to build and validate the models.","marker":"[43]"},{"why":"Defines the optB88 van der Waals exchange functional that corrects the DFT calculations for dispersion interactions between the 2D layer and the bulk surface.","marker":"[40]"},{"why":"Supplies the compositional and structural descriptors used as model inputs for the machine-learning pipeline.","marker":"[48]"},{"why":"Documents experimental growth outcomes for TMDCs on different metal substrates, which the paper uses to check that bulk-dominated stability trends match reality.","marker":"[51]"}],"fun_headline_variants":["Bulk properties, not 2D layer, decide Janus heterostructure stability","Machine learning pins bulk dominance in 2D-bulk interfaces","Bulk electronegativity and surface energy steer Janus 2D-bulk binding","Bulk side sets the energy in Janus 2D-bulk stacks","For 2D-bulk heterostructures, bulk wins over 2D"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that bulk properties dominate stability rests on a dataset in which the 2D materials vary little (electronegativity 1.98–2.59) while the bulk set varies widely (0.95–2.54), so the bulk features have more variance to explain; a wider range of 2D materials could change the ranking.","fun_headline_variants_meta":{"raw":{"variants":["Bulk properties, not 2D layer, decide Janus heterostructure stability","Machine learning pins bulk dominance in 2D-bulk interfaces","Bulk electronegativity and surface energy steer Janus 2D-bulk binding","Bulk side sets the energy in Janus 2D-bulk stacks","For 2D-bulk heterostructures, bulk wins over 2D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1644,"prompt_tokens":1111,"completion_tokens":533,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":727,"completion_tokens_details":{"reasoning_tokens":430}},"tokens_in":727,"tokens_out":533,"duration_ms":5439,"temperature":1.0,"reasoning_tokens":430,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T22:07:49.062432+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Extend the study to non-Janus 2D materials with electronegativities below 1.5 or above 3.0 on the same 19 substrates, recompute binding energies, and retrain the random forest; if 2D electronegativity or the 2D–bulk difference enters the top features, the bulk-dominance claim fails for the broader class.","supporting_citations":[{"cited_title":"S.; Hinsche, N","cited_arxiv_id":null,"evidence_quote":"Supplies the 51 Janus 2D structures and their computed properties from a public 2D materials database; this is the source of the 2D side of the data set."},{"cited_title":"M.; Singh, A","cited_arxiv_id":null,"evidence_quote":"Provides the high-throughput heterostructure construction framework and the DFT parameters and settings used to build and relax every 2D-bulk pair."},{"cited_title":"P.; Richards, W","cited_arxiv_id":null,"evidence_quote":"Supplies the bulk elemental structures and stability data along with analysis tools used to prepare the substrate set."},{"cited_title":"Scikit-learn: Machine Learning in Python","cited_arxiv_id":null,"evidence_quote":"Provides the random-forest regression implementation and the cross-validation and feature-selection routines used to build and validate the models."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the optB88 van der Waals exchange functional that corrects the DFT calculations for dispersion interactions between the 2D layer and the bulk surface."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the compositional and structural descriptors used as model inputs for the machine-learning pipeline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents experimental growth outcomes for TMDCs on different metal substrates, which the paper uses to check that bulk-dominated stability trends match reality."}],"review_version":1}