{"id":"ec88f886-b7f5-4f8f-bbaf-f2bf25a8dc87","arxiv_id":"2509.11231","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A Moment Tensor Potential trained on DFT data reproduces elastic trends, chemical short-range ordering, and stacking fault energies for CoCrNi and CoCrFeNi, with damped CSRO magnitudes and a partially circular CSRO validation.","lead":"This paper builds machine-learned interatomic potentials for CoCrNi and CoCrFeNi and reports they match quantum-mechanical calculations for elasticity, atomic ordering, and stacking faults. It matters because such potentials could unlock larger simulations than DFT permits, though part of the claimed agreement comes from data used in training.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CSRO validation may be circular: the paper states the MTP was trained on the DFT-MC trajectories it is validated against (Section II.D), so the claimed reproduction of DFT CSRO signs could reflect memorization rather than transferable accuracy.","rationale":"The reader identified the spin-polarization encoding as the weakest assumption, but also noted the CSRO comparison is partially circular. I agree the circularity is a more direct and textually supported threat: the paper explicitly states the MTP was trained on DFT-MC trajectories, which are the same kind of data used for validation. If that statement is accurate, the CSRO sign agreement carries little evidentiary weight. This does not necessarily invalidate the potential's other demonstrated strengths (force RMSE, elastic trends), so a conditional verdict remains appropriate—but the condition should be a re-validation on independent CSRO data. The spin concern is real but less decisive, since the MTP could still capture effective magnetic interactions through fitted energies and forces even without explicit spin variables; the paper itself presents the underestimated magnitude as a plausible consequence of missing spin degrees of freedom. My recommendation of CONDITIONAL matches the reader's verdict, but for a different primary reason, hence 'partial' agreement.","tokens_in":17976,"tokens_out":5446,"duration_ms":57817,"concrete_test":"Download the training database from the cited GitLab repository (https://gitlab.com/mtp_potentials/cocrfeni) and identify any configurations that originate from the DFT-MC simulations of Tamm et al. [68] (e.g., by matching supercell size, composition, and MC swap history/provenance). If such overlap exists, retrain the MTP after removing those frames, rerun the 500 K MC/MD CSRO calculation, and compare the resulting Warren-Cowley parameters to [68]. If the sign pattern persists, the concern is resolved; if it changes, the CSRO claim is in-sample and needs external validation.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"In Section II.D, the authors write that 'The MTP, trained on DFT–MC trajectories, again matches the DFT signs across all pairs' when comparing Warren–Cowley parameters to the DFT-MC results of Tamm et al. [68]. This sentence implies the validation set for CSRO overlaps with the training data, making the sign agreement an in-sample result. The Methods section (IV.A) describes a database of unary/binary/ternary/quaternary perturbed structures and does not list MC trajectories, but if the statement in II.D is accurate, the database used for the MTP included configurations sampled by the same DFT-MC procedure used as the CSRO reference. In that case, the central claim that the MTP 'captures DFT-reported CSRO features' is not independently established: an ML potential trained on those configurations is expected to reproduce their signs. The acknowledged underestimation of CSRO magnitudes is then a second-order issue; it does not rescue the validity of the comparison. This concern directly affects the headline result and is distinct from the reader's spin-polarization concern, although both point to the fragility of the CSRO evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript develops Moment Tensor Potentials for the equiatomic medium-entropy alloys CoCrFeNi and CoCrNi, trained on a spin-polarized DFT database spanning unary, binary, ternary, and quaternary structures (Methods IV.A). The authors validate forces (II.A), elastic constants against DFT, EAM, MEAM, and experiments (II.B, Table I), compositional trends of bulk and shear moduli for non-equiatomic compositions (II.C), chemical short-range order via hybrid MC/MD Warren–Cowley parameters (II.D, Table IV), and stacking fault energies (II.E, Table V, Figs. 6–7). The paper claims near-DFT accuracy, reproduction of DFT-reported CSRO signs, and ISF energies consistent with DFT.","tokens_in":18274,"tokens_out":9532,"duration_ms":124457,"significance":"Machine-learned interatomic potentials that are accurate and efficient for CoCrFeNi and CoCrNi would be a useful community resource. The paper has several concrete strengths: the force validation is quantitatively strong (RMSE ~0.15–0.20 eV/Å, R²~0.996); the non-equiatomic composition trends in bulk and shear moduli are a non-trivial transferability test; the chemistry-resolved stacking-fault trends are physically informative; and the potential/training database is promised on a public GitLab page. If the CSRO comparison can be made out-of-sample and the accuracy claims are appropriately calibrated, the work would be a solid contribution. At present, however, the central CSRO claim is compromised by an apparent overlap between training and validation data, and several headline statements overstate the reported quantitative agreement.","major_comments":[{"comment":"The CSRO validation is potentially circular. The text states, \"The MTP, trained on DFT–MC trajectories, again matches the DFT signs across all pairs,\" referring to the DFT-MC results of Tamm et al. [68]. If DFT-MC trajectories were included in the training database, then matching the signs of the same DFT-MC procedure is an in-sample result and does not establish transferability. Methods IV.A lists unary/binary/ternary/quaternary perturbed structures but does not mention MC trajectories, creating an internal contradiction. If the MC trajectories were not part of the training set, the sentence is inaccurate; if they were, the comparison is invalid for the stated purpose. Please clarify the exact composition of the training database and provide an out-of-sample test of CSRO, e.g., by validating against independently generated DFT-MC configurations not used in training.","section":"Section II.D, Table IV"},{"comment":"The claim of \"near-DFT accuracy\" for elastic constants is not supported by the table. For CoCrFeNi, the MTP (SQS) values deviate from DFT by C11 −10.1%, C12 −14.8%, C44 −21.4%, and B −21.9%. For CoCrNi, MTP (SQS) gives C12 −18.1% and C44 −18.5%; even the MTP (SRO) values show C12 −20.3% and C44 −12.9% for CoCrNi. These are not <5% errors, and the EAM/MEAM values are sometimes closer to DFT (e.g., EAM C12 for CoCrFeNi is +2.5%). The abstract and conclusion state that the MTP \"accurately predicts elastic properties\" and \"quantitatively reproduces the lattice constants and elastic constants... achieving near-DFT accuracy.\" This overstates the evidence. Please either report deviations explicitly as a limitation, distinguish SQS vs SRO as the relevant benchmark, or revise the wording.","section":"Table I and Section II.B"},{"comment":"The title and abstract claim DFT accuracy for stacking fault energy, but the quantitative agreement is partial. For CoCrNi, MTP gives γ_ISF ≈ 53.88 ± 5 mJ/m², while the cited DFT value is ~80 mJ/m² [69], a ~33% underestimate; the MTP spread (roughly 40–65 mJ/m²) is also much narrower than the DFT-observed distribution. For CoCrFeNi, MTP gives 35.56 mJ/m² against a DFT range of 17–34 mJ/m², so it matches only the upper bound. The text acknowledges some of this, but the abstract's phrase \"consistent with DFT predictions\" and the title's \"achieving DFT accuracy in ... stacking fault energy\" are not supported. Please reframe these as qualitative/trend-level agreement or provide a more systematic benchmark.","section":"Section II.E, Fig. 6, Table V"},{"comment":"The sign convention for the Warren–Cowley parameter is inconsistent with the physical interpretation. The paper defines α^n_ij = (p^n_ij − c_j)/(δ_ij − c_j). For like pairs (i=j), this gives a positive value when p_ii > c_i, i.e., an excess of like-atom neighbors, which is conventionally clustering/attraction between like species, not \"repulsion.\" Yet the text repeatedly describes positive Cr–Cr and Fe–Fe values as \"strong repulsion\" (e.g., Table IV caption and Fig. 4 discussion). If the reference [68] used the alternative standard definition α_ij = 1 − p_ij/c_j, then positive same-species values would indicate repulsion, but the definitions would not be equivalent and the sign comparisons in Table IV would be invalid. Please state the convention used in both the MTP calculations and the reference data, and correct the terminology accordingly.","section":"Section II.D, Warren–Cowley parameter definition"}],"minor_comments":[{"comment":"The abstract and introduction claim that energies, forces, and stresses are all validated, but Section II.A reports force validation only. Please add energy/stress error metrics or correct the wording.","section":"Abstract and Section II.A"},{"comment":"The description \"r_min was initially set to 2.0 Å and was adaptively updated during training\" is too vague to be reproducible. Please define the update criterion and the range of r_min values used.","section":"Section IV.B"},{"comment":"In the top-three lists, the compositions and modulus values are helpful, but the tables would benefit from explicit column headers clarifying that compositions are in at.% and from a note on whether MTP and DFT used the same MC/MD relaxation protocol.","section":"Tables II and III"},{"comment":"The GitLab link is a positive step. Please add a versioned release or DOI and state explicitly whether the training database includes the DFT-MC trajectories mentioned in Section II.D, since this is critical for evaluating the CSRO validation.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The CSRO circularity concern is real and should be resolved before acceptance. Reference [68] is co-authored by A. Tamm, a co-author of this manuscript; while co-authorship is not itself a problem, the paper should state explicitly that the reference data used for validation are distinct from the training trajectories. The elastic-constant and stacking-fault discrepancies are not fatal if the claims are softened, but the current abstract and title overstate the agreement. Please also verify that the public GitLab repository contains the full training database, since several claims depend on it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a genuinely useful new artifact — a moment tensor potential for CoCrFeNi and CoCrNi trained on a unary-to-quaternary DFT database, with the potential, data, and sample calculations posted on GitLab. The force validation is solid (RMSEs around 0.15–0.20 eV/Å, good directional agreement), and the non-equiatomic elastic trends are a real plus: the MTP ranks low- and high-modulus compositions in line with DFT despite being trained only at equiatomic CoCrFeNi. The stacking-fault chemistry trends (Co-rich planes lower ISF, Cr/Fe-rich planes raise it) are plausible and consistent with prior DFT. So there is a real contribution here.\n\nBut there are two serious soft spots. First, the CSRO validation is circular. Section II.D explicitly says “The MTP, trained on DFT–MC trajectories, again matches the DFT signs across all pairs.” If the MTP was trained on those trajectories, then reproducing their signs is expected; it is not evidence of transferable CSRO accuracy. The Methods section (IV.A) does not list MC trajectories in the database, which makes the sentence confusing — either the database included them and the Methods omission is a lapse, or the sentence is imprecise. Either way, the paper needs to clarify and, ideally, validate on held-out MC structures or an independent reference. The fact that reference [68] is co-authored by one of this paper’s co-authors doesn’t help the independence impression.\n\nSecond, the abstract overclaims. For CoCrNi, the MTP gives ISF ≈54 mJ/m² while the cited DFT value is ≈80 mJ/m² — that is not “consistent with DFT predictions” in the usual sense, and the broad DFT distribution (~−6 to 288 mJ/m²) makes the comparison even murkier. Elastic constants deviate by up to 20% from DFT for C44 in CoCrFeNi (136 vs 173 GPa), so calling the agreement “near-DFT accuracy” is too strong. The paper’s own text acknowledges the CSRO magnitudes are damped, which is honest, but it also reveals a deeper limitation: magnetic effects are only implicitly encoded, and the authors themselves attribute the underestimated CSRO to missing spin polarization. That is a plausible interpretation, but it is not tested.\n\nNone of this kills the paper. The potential is likely to be useful for large-scale MD studies, and the force validation, elastic ordering, and SFE chemistry trends stand on their own. But the central CSRO claim, as written, is not established, and the abstract needs recalibration.\n\nMy recommendation: send it to peer review, but ask for a revision that fixes the CSRO validation (or clearly separates in-sample from independent checks), clarifies the database contents, and tempers the abstract. With those changes, this would be a solid contribution worth citing.","headline":"Useful new MTP for CoCrFeNi/CoCrNi, but the CSRO validation is in-sample by the paper's own admission, and the abstract oversells the agreement.","tokens_in":18795,"tokens_out":2349,"would_cite":true,"duration_ms":29953,"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":"A machine-learned interatomic potential brings DFT-level accuracy to two medium-entropy alloys, capturing chemical short-range order and stacking fault energies at a fraction of DFT's computational cost.","keywords":["medium-entropy alloys","moment tensor potential","chemical short-range order","stacking fault energy","machine learning interatomic potential","CoCrFeNi","CoCrNi","elastic properties"],"falsifier":"A decisive test would compute Warren–Cowley parameters for CoCrNi at 500 K from MTP-based lattice Monte Carlo using a supercell size and MC schedule identical to the DFT benchmark; if the MTP values remain damped when system sizes are matched, the missing explicit spin polarization—not sampling—causes the discrepancy.","tokens_in":17876,"feed_emoji":"⚛️","tokens_out":5989,"duration_ms":61239,"temperature":0.7,"pith_summary":"The paper sets out to show that a moment tensor potential (MTP), a machine-learned interatomic potential, can close the accuracy gap between density functional theory (DFT) and classical empirical potentials for the FCC medium-entropy alloys CoCrFeNi and CoCrNi. The authors train the MTP on a DFT database spanning unary, binary, ternary, and quaternary structures, including spin-polarized energies, forces, and stresses, then test it on elastic properties, chemical short-range ordering (CSRO), and stacking fault energetics. They report near-DFT force errors, elastic constants in near-quantitative agreement with experiment, CSRO sign patterns matching DFT, and intrinsic stacking fault energies near 54 mJ/m² for CoCrNi and 36 mJ/m² for CoCrFeNi. A sympathetic reader would care because this enables molecular dynamics simulations of defects and mechanical behavior at scales DFT cannot reach, while resolving chemistry effects that classical potentials miss. The paper itself states the main caveat: CSRO magnitudes are damped relative to DFT, attributed to the lack of explicit spin polarization and to differences in system size and sampling.","feed_headline":"Machine-learned potential matches DFT for two medium-entropy alloys","feed_subtitle":"Moment tensor potential captures chemical ordering and stacking fault energies in CoCrFeNi and CoCrNi at a fraction of DFT cost.","key_machinery":"The central machinery is the moment tensor potential (MTP), a machine-learned interatomic potential that represents each atom's local environment by moment tensors—products of radial functions and tensor products of neighbor displacement vectors—up to a chosen expansion level (here level 22). The potential is trained on energies, forces, and stresses from spin-polarized DFT for unary, binary, ternary, and quaternary configurations. The load-bearing device is the implicit encoding of magnetic interactions: the MTP has no spin variables, yet the paper asserts that because it is fitted to spin-polarized DFT energies and forces, it inherits magnetic effects. This implicit encoding is what lets t","core_discovery":"On its own terms, the paper demonstrates that a moment tensor potential fitted to spin-polarized DFT data—without explicit spin degrees of freedom—reproduces chemical-ordering preferences and stacking-fault energies in CoCrFeNi and CoCrNi. Force RMSEs are 0.153 eV/Å (CoCrNi) and 0.203 eV/Å (CoCrFeNi); elastic constants lie within a few percent of DFT; Warren–Cowley parameters have correct signs for every nearest-neighbor pair; ISF energies are 53.9 and 35.6 mJ/m². The potential extrapolates to non-equiatomic compositions, matching DFT's bulk and shear modulus rankings, and captures fault-plane chemistry. The claimed net result is a transferable model with DFT-level fidelity at molecular-dyna","pith_inferences":["If the size/sampling explanation for the damped CSRO magnitudes is correct, re-running the MTP Monte Carlo with DFT-sized supercells and matching MC schedules should raise the Warren–Cowley magnitudes toward the DFT values; if they remain low, implicit spin encoding is the limiting factor.","The same training recipe could be stress-tested on magnetically complex alloys containing manganese, such as CoCrFeMnNi, where local moments are stronger and explicit spin degrees of freedom may be harder to sidestep.","The sign-structure accuracy suggests the MTP could be used to train coarser models—for example, effective pair interactions or cluster expansions—that explicitly encode CSRO tendencies at larger length scales."],"forward_implications":["If the MTP transfers as claimed, large-scale molecular dynamics of dislocation glide, twinning, and radiation damage in CoCrFeNi and CoCrNi can use near-DFT accuracy instead of empirical potentials that misrepresent CSRO.","The potential can screen non-equiatomic alloy compositions for elastic properties cheaply, identifying promising Fe/Co/Ni/Cr ratios before experiment.","Capturing chemistry-dependent stacking fault energies means the MTP can predict how local segregation changes deformation mechanisms such as twinning and phase transformation.","The trained potential and training database are publicly available, so other groups can validate and extend the approach to related high-entropy alloys."],"fun_headline_variants":["Moment tensor potential hits DFT-grade accuracy for two alloys","ML potential nails short-range ordering and fault energies","DFT-level predictions without DFT price tag for CoCrFeNi and CoCrNi","MTP reproduces DFT energies for chemical order and faults"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The assumption that carries the argument is that spin-polarized DFT energies and forces encode magnetic interactions well enough that a potential without explicit spin variables can still reproduce the chemistry that magnetism drives; if that encoding fails for Cr-rich or manganese-containing local environments, the CSRO and stacking-fault predictions lose their physical base.","fun_headline_variants_meta":{"raw":{"variants":["Moment tensor potential hits DFT-grade accuracy for two alloys","ML potential nails short-range ordering and fault energies","DFT-level predictions without DFT price tag for CoCrFeNi and CoCrNi","MTP reproduces DFT energies for chemical order and faults"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00032,"raw_usage":{"total_tokens":1705,"prompt_tokens":874,"completion_tokens":831,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":760}},"tokens_in":618,"tokens_out":831,"duration_ms":10072,"temperature":1.0,"reasoning_tokens":760,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T16:51:20.528988+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test would compute Warren–Cowley parameters for CoCrNi at 500 K from MTP-based lattice Monte Carlo using a supercell size and MC schedule identical to the DFT benchmark; if the MTP values remain damped when system sizes are matched, the missing explicit spin polarization—not sampling—causes the discrepancy.","supporting_citations":[],"review_version":1}