{"id":"de1eeb8e-ce0b-495f-a1a5-b9984bdd0260","arxiv_id":"2507.18873","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A moment tensor potential trained on DFT data predicts Ti-N formation energies with test RMSE 6.8 meV/atom and suggests intermediate Ti-N phases across the full composition range are thermodynamically stable.","lead":"Researchers built a machine-learned interatomic potential for the titanium-nitrogen system that reproduces density functional theory formation energies to within a few meV per atom. The potential could enable faster simulations of TiN coatings, diffusion barriers, and other Ti-N materials.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The stability claim for intermediate N/Ti ratios rests on MTP extrapolation to structures never checked against DFT; the two held-out tests do not cover the insertion/removal structures used in Section 3.7.","rationale":"The MTP development is reasonable as far as it goes: it is trained on DFT data including strained and finite-temperature configurations, and it is tested on Ti6N5 and Ti-0.2N with formation-energy errors in the tens of meV/atom and elastic-constant median errors around 5%. That supports using the potential for properties near the training distribution. The stress-test issue is specifically the leap from good performance on two held-out ordered/vacancy/solid-solution structures to the claim that all N/Ti ratios are thermodynamically stable. Section 3.7 uses only MTP0.75-0.25-0.25 to relax hundreds of generated structures and assign formation energies; there is no DFT check on any of them. The held-out errors, up to about 7.6 meV/atom at the KDE peak, are comparable to the claimed 10 meV/atom hull margin, so error bars alone could change the hull membership. The heuristic generation protocol, farthest-octahedral insertion and iterative single removal, also means the lowest-energy structure found is not guaranteed to be the ground state at that composition; however, the existence claim is less sensitive to search completeness than to MTP accuracy. A small DFT validation set across the composition range would settle the question. If the DFT checks confirm the MTP hull, the central claim stands; if not, the claim should be weakened to 'MTP predicts candidate stable structures'. Therefore the reader's CONDITIONAL verdict is appropriate and no adjustment is needed. The absence of released potential and training data is a reproducibility concern, but it is secondary to the scientific validation gap.","tokens_in":15006,"tokens_out":5200,"duration_ms":51141,"concrete_test":"Select the lowest-energy MTP-relaxed structure from Section 3.7 at five or more N/Ti ratios not in the training set (for example, 0.125, 0.25, 0.375, 0.625, 0.875). Fully relax each with VASP PBE using the same 520 eV cutoff and 0.03 Angstrom^-1 k-point mesh as Section 3.1, and compute formation energies relative to hcp Ti and N2. Compare to the MTP values and check whether DFT relaxation preserves the MTP-generated N ordering. If any DFT formation energy lies more than 10 meV/atom above the MTP value, or if the structure reconstructs, the universal-stability claim is unsupported and Section 3.7 needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.7 constructs the 0 K hull for N/Ti between 0 and 1 by generating structures through nitrogen insertion and removal and relaxing every structure with MTP0.75-0.25-0.25. The paper's central claim that all intermediate stoichiometries can be thermodynamically stable therefore depends on the potential being accurate for these generated structures. The only evidence of transferability is two held-out cases, Ti6N5 and Ti-0.2N, with formation-energy peak errors of 5.9 and 7.6 meV/atom (Section 3.4.2). These are chemically related to the training set (ordered N-vacancy derivatives of TiN and an interstitial hcp solid solution), whereas Section 3.7 explores arbitrary configurations, such as random N insertion at farthest octahedral voids and iterative single-N removals, whose local environments are not represented in the test set. No DFT calculation is reported for any of the newly predicted structures. Since the stability claim is a 0 K convex-hull statement and the predicted structures are claimed to be within about 10 meV/atom of the hull, errors of the size seen in held-out tests are large enough to move a structure from on-hull to off-hull and invert the conclusion. The potential itself appears reasonable on the tested compounds, but it has not been shown to support the extrapolation that carries the headline conclusion.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a moment tensor potential (MTP) for the Ti-N system, trained on DFT data for Ti, TiN, Ti2N, Ti3N2, Ti4N3, and the solid solution Ti-0.14N, with Ti6N5 and Ti-0.2N held out for testing. Three weighting schemes are compared; the authors select MTP0.75-0.25-0.25 as the most accurate based on formation-energy error distributions and elastic-constant errors. The selected potential is then used to relax and compute formation energies of structures generated by inserting N at octahedral voids and removing N from ordered compounds, leading to the central claim that structures with N/Ti ratios ranging from 0 to 1 can be thermodynamically stable, with a maximum deviation of 10 meV/atom from the 0 K convex hull.","tokens_in":15238,"tokens_out":3311,"duration_ms":33953,"significance":"If fully validated, this work would provide a fast and accurate interatomic potential for a technically important system and a broadly applicable approach to training-set selection for MTPs. The reported held-out formation-energy errors (RMSE 6.8 meV/atom; peak errors 5.9 and 7.6 meV/atom for the two test structures) and elastic-constant median error (~4.8%) are respectable for an MTP, and the systematic comparison against MEAM and M3GNet is useful. However, the headline conclusion about the thermodynamic stability of new intermediate phases rests entirely on MTP-extrapolated energies with no DFT confirmation for the predicted structures; as presented, that conclusion is not yet established.","major_comments":[{"comment":"The central claim that 'structures with N/Ti ratios ranging from 0 to 1 can be thermodynamically stable' is based solely on formation energies computed with the MTP0.75-0.25-0.25 potential for structures generated by arbitrary N insertion/removal. None of these predicted structures is checked against DFT. The only held-out evidence of transferability is Ti6N5 and Ti-0.2N (Section 3.4.2), which are chemically close to the training set (an ordered N-vacancy derivative of TiN and an interstitial hcp solid solution). The Section 3.7 generated structures, such as random N placement at farthest octahedral voids and iterative single-N removals, include local environments not represented in the test set. To support the hull claim, representative predicted structures at several intermediate compositions should be relaxed with DFT and their formation energies compared with the MTP values.","section":"Section 3.7"},{"comment":"The claimed 10 meV/atom maximum deviation from the convex hull is of the same order as the potential's held-out errors (peak absolute errors 5.9 and 7.6 meV/atom in Section 3.4.2). An energy error of this magnitude is sufficient to move a structure from on-hull to off-hull and to invert the relative stability of competing configurations. The paper should quantify the uncertainty of the predicted hull points, for example by propagating the observed test-set error distribution into the hull construction or by DFT-validating the near-hull structures.","section":"Section 3.7 / Abstract"},{"comment":"The KDE/FWHM values reported for MTP1-0-0 on Ti-0.14N (FWHM 2.091E+018, peak position 4.063E+018 meV/atom) and on Ti-0.2N (FWHM 1.8721E+018, peak position 8.2722E+018 meV/atom) are unphysical and indicate either numerical overflow, outliers in the error data, or a computational bug in the KDE implementation. These entries are reproduced in the text as 'orders of 10^18' without explanation. Such artifacts must be diagnosed and corrected, as they undermine confidence in the statistical analysis that is used to select the final potential.","section":"Section 3.4.2, Tables 3 and 4"}],"minor_comments":[{"comment":"The phrase 'maximum deviation of 10 meV/atom from the convex hull plot of formation energy 0K' is ambiguous: it should specify whether the deviation is measured from the DFT hull, from the MTP hull, or from the hull constructed using the predicted structures themselves.","section":"Abstract and Section 4"},{"comment":"The bandwidth formula uses sigma, but the standard deviation of the error distribution is not defined in the text; please define sigma precisely.","section":"Section 3.4.2, Eq. (2)-(4)"},{"comment":"The sentence listing supercells ('Ti2N ,Ti3N2, Ti4N3, Ti6N5 supercell containing 36, 72, 42, 44 and 64 atoms') has five atom counts for five structures but names only four; please correct the list and clarify which structure corresponds to which count.","section":"Section 3.5"},{"comment":"There are numerous typographical and grammatical errors, including 'Ti-N material system have', 'solutions likeTi', 'MTP0.75−0.25−0.25' written without spaces, and 'Nose Hoover thermostat' for Nosé-Hoover. A careful proofreading pass is needed.","section":"Throughout"},{"comment":"The radial function f_mu appearing in the moment tensor definition is not defined; please specify its form and how the 421 basis functions and 2621 descriptors are derived from it.","section":"Section 2.1, Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"The paper's central new-phase stability claim is not yet supported because it relies entirely on the fitted potential without DFT confirmation of the predicted structures. The reported KDE overflow values (1e+18 meV/atom) are also a serious red flag that should be addressed before publication. I would encourage the authors to validate a representative subset of the Section 3.7 structures with DFT and to deposit the MTP potential and training data for reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the fitted MTP is a solid piece of work for the known Ti-N compounds and solid solutions, but the paper's central claim about thermodynamic stability across the whole N/Ti range goes beyond what the validation supports. The two held-out tests (Ti6N5, Ti-0.2N) are too few and too close to the training set to license predictions for the insertion/removal structures in Section 3.7.\n\nWhat is actually good: the selection of training data using structural similarity (ordered N-vacancy derivatives of TiN, hcp solid solutions) is sensible and worked. The held-out RMSE of 6.8 meV/atom for formation energy and median elastic constant errors around 5% are genuine, and the comparison shows MTP clearly beats MEAM and M3GNet for this system. The elastic-constant benchmark and the KDE error analysis are carefully done. If you need a Ti-N potential for atomistic simulations of known phases, this one is likely a real improvement.\n\nWhere it goes soft: Section 3.7 constructs the 0 K hull for N/Ti between 0 and 1 by generating structures via N insertion and removal, relaxing every structure with MTP0.75-0.25-0.25, and then identifying the lowest-energy one. No DFT is reported for any of these generated structures. The only evidence of transferability is a two-structure test set, and the generated configurations (random insertion at farthest octahedral voids, iterative single-N removals) explore local environments that are not really represented in those tests. With held-out peak errors of 5.9 and 7.6 meV/atom, and the paper claiming hull deviations up to 10 meV/atom, the error bars are large enough to move structures on and off the hull. So the headline conclusion is not established. Also, no potential file or training set is released, which makes independent checking impossible; that would be an easy fix and should be required.\n\nNet: the potential development part deserves peer review and is useful to the community. The stability prediction part needs DFT verification on a sample of the predicted structures before it can be taken seriously. If I were the editor I'd send it out, with the expectation of major revision.","headline":"A fitted MTP for Ti-N that is accurate on known phases, but the claim that all intermediate stoichiometries are thermodynamically stable is an unvalidated extrapolation from the potential.","tokens_in":15824,"tokens_out":3958,"would_cite":false,"duration_ms":36816,"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":"This paper develops a moment tensor potential that predicts formation energies, elastic constants, and thermodynamic stability across the entire Ti-N composition range, including previously uncharacterized intermediate structures.","keywords":["Ti-N system","moment tensor potential","interatomic potential","formation energy","convex hull","elastic constants","machine learning","thermodynamic stability"],"falsifier":"Take a random sample of the intermediate structures generated in Section 3.7, for instance ten compositions along the N/Ti range covered by nitrogen insertion into pure Ti and nitrogen removal from Ti2N, and compute their formation energies with DFT at the same level used to train the potential. If more than a few of these energies deviate from the MTP predictions by more than about 10 meV/atom, the claim that structures across the full 0-to-1 range are thermodynamically stable would not survive.","tokens_in":14766,"feed_emoji":"⚛️","tokens_out":7148,"duration_ms":65897,"temperature":0.7,"pith_summary":"Titanium nitride is not a single compound but a family: ordered phases such as Ti2N, Ti3N2, Ti4N3, and Ti6N5, together with nitrogen solid solutions in titanium, span a composition range from pure Ti to TiN. This paper tries to establish that one machine-learned moment tensor potential can capture the energy, mechanical response, and thermodynamic stability of the whole family, not just the handful of known compounds. The authors build a training set by exploiting the structural relationship among the phases — most are rock-salt TiN with ordered nitrogen vacancies, while the solid solutions are nitrogen in octahedral voids of hcp Ti. On held-out structures the potential reaches formation-energy errors of a few meV per atom and elastic-constant errors with a median near 5%. Using this potential, they predict that intermediate structures with N/Ti ratios from 0 to 1 can be thermodynamically stable, within about 10 meV/atom of the 0 K convex hull.","feed_headline":"Machine-learned potential maps Ti-N stability from pure Ti to TiN","feed_subtitle":"Formation energies land within a few meV per atom of DFT, opening the whole composition range to fast simulation.","key_machinery":"The central object is the moment tensor potential, a machine-learned interatomic potential that represents each atom's local environment by moment tensors built from radial functions and tensor products of neighbor position vectors, with a cut-off radius of 7 Å and 2621 moment-tensor descriptors. The specific variant selected, MTP0.75-0.25-0.25, weights energy, force, and stress terms in the loss function as 0.75, 0.25, and 0.25; the paper finds this balanced weighting is what makes energy predictions accurate for structures outside the training set. The argument is carried by a training-set design that treats Ti3N2, Ti4N3, and Ti6N5 as ordered nitrogen-vacancy derivatives of rock-salt TiN and treats solid solutions as nitrogen occupying octahedral voids in hcp Ti, so the same local environments recur across compositions. A convex hull of formation energy versus composition is the device that turns the potential's predictions into a statement about thermodynamic stability.","core_discovery":"The central claim is that a single MTP interatomic potential, trained on Ti, Ti2N, Ti3N2, Ti4N3, TiN, and one solid solution (Ti-0.14N) and validated on Ti6N5 and Ti-0.2N, reliably describes formation energies, elastic constants, and thermodynamic stability across the full Ti-N system. The chosen potential, MTP0.75-0.25-0.25, reproduces DFT formation energies with a root-mean-square error of 2.1 meV/atom on training data and 6.8 meV/atom on the two test structures; the peak of the absolute-error distribution lies at 3.8 meV/atom for training-set systems and 7.6 meV/atom for unseen systems. Elastic constants predicted by the potential show a cumulative median error of 4.83% and an interquartile range of 6.12%, better than the two alternative MTP weightings considered. The authors then use the potential to generate intermediate structures by inserting nitrogen into octahedral voids and removing nitrogen from known phases, relaxing every candidate with the potential, and find that structures across the entire N/Ti range from 0 to 1 fall on or near the 0 K convex hull, with a maximum reported deviation of about 10 meV/atom.","pith_inferences":["Going beyond the paper, the same iterative nitrogen insertion and removal protocol could map off-stoichiometric stability in other interstitial metal-nitride or metal-hydride systems.","If the predicted intermediate phases are confirmed by DFT or experiment, the potential could enable interface-scale simulations of Ti/TiN diffusion barriers, where composition gradients pass through these intermediate stoichiometries.","The validation set is small (two held-out structures), so a natural extension is to test the potential against a broader random sample of off-stoichiometric configurations before relying on the full predicted phase diagram."],"forward_implications":["The potential can compute formation energies of arbitrary Ti-N compositions at near-DFT accuracy, making full-composition convex-hull screening practical without new DFT runs.","Elastic constants across Ti-N compounds and solid solutions can be predicted with a median error around 5%, supporting estimates of bulk, shear, and Young's moduli from the same potential.","Molecular dynamics simulations of Ti-N compounds remain stable from 10 K to 300 K, so finite-temperature behavior can be studied with the same potential.","The success of including force and stress weights in training suggests that energy-only fitted potentials are insufficient for transferable predictions across a composition landscape.","New structures with N/Ti between 0 and 1 are predicted thermodynamically stable within 10 meV/atom of the 0 K convex hull, identifying specific intermediate stoichiometries as synthesis targets."],"supporting_citations":[{"why":"First-principles study establishing the structural relationship among titanium nitride phases and predicting Ti6N5, which shapes the training-set selection and provides the held-out test structure.","marker":"[9]"},{"why":"Existing modified embedded-atom method potential for Ti-C and Ti-N that this work compares against and aims to surpass in coverage of intermediate compounds and solid solutions.","marker":"[10]"},{"why":"Moment tensor potential application to Ni-Al alloys demonstrating accuracy over traditional potentials, motivating the choice of MTP for Ti-N.","marker":"[15]"},{"why":"Original moment tensor potential formulation that defines the representation used to build the potential.","marker":"[17]"},{"why":"First-principles study of Ti4N3 and Ti6N5 under pressure that supplies reference behavior for the nitrogen-vacancy phases used in training and testing.","marker":"[30]"},{"why":"Density functional theory study of Ti2N formation by nitrogen intercalation in titanium, directly relevant to the insertion protocol used to generate new structures.","marker":"[31]"},{"why":"Maxvol algorithm used to reduce the molecular-dynamics and strained configurations to a representative training subset.","marker":"[32]"},{"why":"Pretrained graph neural network potential used as a baseline comparison on the convex hull, showing the gap the trained MTP closes.","marker":"[36]"}],"fun_headline_variants":["Single MTP potential nails Ti-N energetics and stability","Machine-learned potential predicts all Ti-N stable phases","Ti-N formation energies within 6.8 meV of DFT across range","One potential, full Ti-N stability map with meV accuracy","MTP potential predicts Ti-N phases with DFT-level accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole map of stable intermediate phases rests on the assumption that a potential trained on a small set of known compounds and two solid solutions stays accurate for the many off-stoichiometric structures it was never tested against.","fun_headline_variants_meta":{"raw":{"variants":["Single MTP potential nails Ti-N energetics and stability","Machine-learned potential predicts all Ti-N stable phases","Ti-N formation energies within 6.8 meV of DFT across range","One potential, full Ti-N stability map with meV accuracy","MTP potential predicts Ti-N phases with DFT-level accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001035,"raw_usage":{"total_tokens":4441,"prompt_tokens":1116,"completion_tokens":3325,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":732,"completion_tokens_details":{"reasoning_tokens":3240}},"tokens_in":732,"tokens_out":3325,"duration_ms":24551,"temperature":1.0,"reasoning_tokens":3240,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:05:40.150854+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of the intermediate structures generated in Section 3.7, for instance ten compositions along the N/Ti range covered by nitrogen insertion into pure Ti and nitrogen removal from Ti2N, and compute their formation energies with DFT at the same level used to train the potential. If more than a few of these energies deviate from the MTP predictions by more than about 10 meV/atom, the claim that structures across the full 0-to-1 range are thermodynamically stable would not survive.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"First-principles study establishing the structural relationship among titanium nitride phases and predicting Ti6N5, which shapes the training-set selection and provides the held-out test structure."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Moment tensor potential application to Ni-Al alloys demonstrating accuracy over traditional potentials, motivating the choice of MTP for Ti-N."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"First-principles study of Ti4N3 and Ti6N5 under pressure that supplies reference behavior for the nitrogen-vacancy phases used in training and testing."},{"cited_title":"Varalakshmi, S","cited_arxiv_id":null,"evidence_quote":"Density functional theory study of Ti2N formation by nitrogen intercalation in titanium, directly relevant to the insertion protocol used to generate new structures."}],"review_version":1}