{"id":"f328af4b-ccca-48ed-abe0-2fdc50d3d777","arxiv_id":"2412.20214","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"An actively trained magnetic Moment Tensor Potential fitted with constrained DFT reproduces elastic, phonon, and thermal properties of paramagnetic B1-CrN.","lead":"Researchers developed an automated active learning protocol for fitting magnetic Moment Tensor Potentials, machine-learned interatomic models that explicitly include atomic magnetic moments. They trained the potential on chromium nitride (CrN) using constrained density functional theory and showed it reproduces the material's elastic, vibrational, and thermal properties in its paramagnetic state.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The most load-bearing assumption is that equal-weight averaging over 50 random collinear spin states (Section III.B) represents the paramagnetic state; the reported convergence test does not rule out finite-temperature short-range spin correlations, and all PM predictions inherit this risk.","rationale":"The reader identified the same load-bearing assumption: the paramagnetic state is represented by an infinite-temperature equal-weight average over 50 random collinear spin configurations. I agree that this is the most critical point because it underpins every paramagnetic property claimed in the abstract and in the validation sections. The paper's own convergence test (Fig. 5) addresses only the number of configurations needed for statistical stability of the average, not the physical representativeness of the infinite-temperature measure. Near the Néel temperature, short-range magnetic correlations can persist and may systematically shift forces, stresses, and phonon frequencies; the reported fitting errors cannot detect such a bias because the training set is generated from the same ensemble. The comparison to DFT[18] is reassuring but not decisive, since that reference itself uses an approximate disorder model. The heat-capacity deviation from 280 to 400 K is an additional warning that the model's temperature dependence is not fully captured. The proposed finite-temperature spin-flip sampling test would settle whether the infinite-T ensemble is adequate; if it passes, the central claim stands. Since the reader's conditional verdict already includes the need for error bars and sensitivity checks, and this concern reinforces those conditions without changing the overall assessment, no adjustment to the verdict is needed.","tokens_in":13495,"tokens_out":5991,"duration_ms":67057,"concrete_test":"Run a finite-temperature spin-flip Monte Carlo (or DLM-MD with spin updates) at T = 300, 400, and 600 K using the trained mMTP energy (or a Heisenberg model fitted to the same cDFT data), and recompute the paramagnetic elastic constants and phonon spectrum as Boltzmann-weighted averages over the sampled spin states. If the resulting C11/C44 and phonon frequencies differ from the equal-weight 50-configuration averages by more than the reported fitting errors (stress error 0.4 GPa, energy error 1.7 meV/atom) or by more than the visible linewidth in Fig. 8, the infinite-T ensemble is not representative and the central claim should be restricted to the infinite-T model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"All paramagnetic-state predictions (Table II PM row, Figs. 8-10) are generated from the ensemble defined in Section III.B: an equal-weight average over 50 randomly disordered collinear magnetic configurations with 50% up / 50% down moments. The authors explicitly state that this averaging 'corresponds to the limit of infinite temperatures.' The convergence check in Fig. 5 only demonstrates that adding more random configurations changes the averaged energy and force for one displaced Cr atom by less than the mMTP fitting errors; it does not establish that the infinite-temperature measure is the physically correct one for paramagnetic CrN at temperatures just above the Néel temperature (280 K), where short-range magnetic correlations are expected to be non-negligible. Because the mMTP is trained on configurations drawn from this same ensemble, a biased ensemble cannot be detected by fitting error or by active-learning extrapolation grades; it enters every reported PM quantity as a systematic offset. The paper provides no comparison against a finite-temperature magnetic sampling (e.g., the Boltzmann-weighted SSA used in ref. [18], or DLM/MC spin-flip sampling), so the agreement with DFT[18] is partly an agreement between two approximate disorder models. The heat-capacity discrepancy in the 280-400 K window (Fig. 10) is acknowledged but not quantified in a way that separates the ensemble limitation from the missing Ortho-CrN phase.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an automated active-learning protocol for fitting magnetic Moment Tensor Potentials (mMTPs) in which training configurations are selected during magnetic-moment equilibration, relaxation, and molecular dynamics, and are subsequently labeled by constrained DFT (cDFT) energies, forces, stresses, and magnetic forces. The method is applied to B1-CrN, and the paramagnetic state is represented by an equal-weight average over 50 randomly disordered collinear magnetic configurations. The authors report elastic constants, phonon spectra, thermal expansion, and specific heat capacity for B1-CrN, and claim that the automatically trained mMTP reproduces mechanical, dynamical, and thermal properties of the paramagnetic state with respect to DFT and experiments. The training set is small (2423 configurations) and is made openly available.","tokens_in":13711,"tokens_out":4106,"duration_ms":44516,"significance":"If the central claim holds, the paper is a useful methodological contribution: it replaces manual training-set construction for magnetic MLIPs with an active-learning loop that exploits cDFT to label non-equilibrium magnetic configurations, and it demonstrates the approach on a material with a complex magnetic order-disorder transition. The open dataset and the explicit fitting to magnetic forces are strengths, and the comparison against independent experimental thermal data gives the paper a falsifiable component. The main significance hinges, however, on whether the equal-weight random-spin ensemble used for the paramagnetic state is physically representative at the temperatures of interest, and on whether the reported agreement with DFT is as quantitative as the text suggests.","major_comments":[{"comment":"The paramagnetic state is defined as an equal-weight average over 50 random collinear spin configurations, which the authors state corresponds to the infinite-temperature limit. The convergence test in Fig. 5 shows that adding more configurations changes the averaged energy and force by less than the mMTP fitting errors, but this only controls statistical sampling error; it does not address whether the infinite-temperature measure is representative of the short-range-correlated paramagnetic state just above the Néel temperature (280 K). Since every paramagnetic prediction in Table II and Figs. 8-10 is generated from this ensemble, a systematic bias in the ensemble would enter all of the central results. I ask the authors to provide a concrete test of ensemble representativeness, for example a comparison with Boltzmann-weighted spin-space averaging as used in Ref. [18], or a sensitivity study over ensembles with different degrees of short-range order, and to state clearly how the reported properties change under that test.","section":"Section III.B"},{"comment":"The text states that the mMTP and DFT elastic constants \"coincide very well,\" but the table shows substantial disagreements for shear constants. For the AFM state, C44 is 158 GPa from mMTP versus 122 GPa (DFT[*]) and 124 GPa (DFT[18]), a deviation of about 28%. For the FM state, C44 is 188 GPa versus 164 and 162 GPa, a deviation of about 15%. Even the paramagnetic C44 deviates by about 10% (160 vs 145 GPa). The claim of very good agreement should be either quantified with error metrics or tempered, and the authors should discuss what these deviations imply for the intended downstream use of the potential in mechanical and dynamical simulations.","section":"Table II"},{"comment":"The claim that the fitted mMTP reproduces thermal properties is supported by specific heat capacity agreement only above about 400 K. The discrepancy between 280 and 400 K is acknowledged and attributed to the absence of Ortho-CrN data, but it is not quantified, and the attribution is not tested. I request a numerical comparison (e.g., average absolute or relative error in the 280-400 K window and above 400 K), and either a demonstration that the discrepancy is dominated by the missing Ortho-CrN phase (for instance, a calculation of the Ortho-CrN/AFM contribution to the heat capacity) or an explicit statement that the thermal-property claim in the abstract and conclusion is limited to the B1 paramagnetic phase above 400 K.","section":"Section III.E, Fig. 10"}],"minor_comments":[{"comment":"The heading contains a typo: \"Magnetic Moment T ensor Potential\" should read \"Magnetic Moment Tensor Potential.\"","section":"Section II.A"},{"comment":"The caption reads \"experimentally [18]\"; it should read \"experimental data from [18]\" or \"experimental values from [18].\"","section":"Figure 9 caption"},{"comment":"The sentence describing perturbations of magnetic moments is ambiguous: \"we repeated the previous step for the magnetic moments by applying perturbations, again, two times, by at most 15% to each equilibrium magnetic moment.\" Please specify whether each of the two repetitions uses the same 15% bound and whether the perturbed moments are independent of the atomic displacements.","section":"Section III.A"},{"comment":"The term \"non-equilibrium magnetic moments\" is used frequently; a brief definition or a reference to the cDFT description would help readers understand that these are magnetic moments that are not the self-consistent ground-state moments for the given atomic configuration.","section":"Section III.A"},{"comment":"The labels DFT[*] and DFT[18] are clear enough, but the caption could state explicitly which magnetic ordering and which computational settings (U value, k-point mesh, cutoff) correspond to DFT[*] so that the comparison is reproducible without going back to Section II.D.","section":"Table II"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the methodology is promising. The main issue is that the central claim about the paramagnetic state rests on an ensemble whose physical representativeness is not validated beyond statistical convergence. The authors' own future-work discussion (DLM-MD and MDMC) implicitly concedes that the current sampling does not account for temperature-dependent spin correlations. I recommend major revision to require either a quantitative ensemble-robustness test or a carefully scope-limited claim, and to address the quantitative overstatement in Table II."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this paper actually delivers on automating training-set construction for magnetic MLIPs, which prior work did manually. The active-learning loop with cDFT (including magnetic forces) is new, and the demonstration on B1-CrN is thorough: FM and AFM elastic constants and phonons match DFT well, and the PM thermal expansion and high-T heat capacity track experiment. The training set is only 2423 configurations, which is small, and the data are openly linked. That is a real contribution for the MLIP and magnetic-materials community.\n\nSoft spots, in proportion. The biggest is the paramagnetic ensemble: equal-weight averaging over 50 random collinear spin states is the infinite-temperature limit. The convergence check in Fig. 5 shows the average stabilizes, but it does not validate that this measure captures short-range correlations just above the Néel temperature. All PM predictions inherit that risk as a possible systematic offset. The paper does not compare against finite-temperature sampling (e.g., SSA or DLM), so the agreement with DFT[18] is partly agreement between two disorder models. That said, the authors are explicit about their assumption and the heat-capacity discrepancy below 400 K is honestly attributed to the missing Ortho-CrN phase. The PM C44 deviation from DFT[18] (160 vs 145 GPa) is a minor but real mismatch.\n\nTwo smaller things. The Hubbard U was chosen to match the experimental lattice parameter, which makes the thermal expansion agreement partly circular; not fatal, but sensitivity to U is untested. And the abstract claims agreement with experiments a bit more broadly than the body justifies.\n\nVerdict: this deserves a serious referee. The conditions are reasonable: report spread/error bars on spin-averaged quantities, test sensitivity to U and key hyperparameters, and soften the abstract to match the demonstrated scope. The core protocol result holds up. I would cite it and bring it to reading group, and I would accept it after revision.","headline":"A solid methods paper that genuinely automates training of magnetic moment tensor potentials with cDFT and active learning, with honest validation on CrN; the main caveat is the unquantified infinite-temperature assumption behind the paramagnetic ensemble.","tokens_in":14380,"tokens_out":839,"would_cite":true,"duration_ms":10401,"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":"An automatically trained magnetic Moment Tensor Potential reproduces elastic, dynamical, and thermal properties of paramagnetic B1-CrN against DFT and experiment.","keywords":["magnetic moment tensor potential","active learning","constrained density functional theory","paramagnetic CrN","elastic constants","phonon spectra","lattice thermal expansion","specific heat capacity"],"falsifier":"Compute the elastic constants, phonon spectra, and thermal expansion of B1-CrN using a paramagnetic ensemble that includes finite-temperature short-range spin correlations (for example, from disordered-local-moment molecular dynamics or Monte Carlo spin flips at realistic temperatures) and compare against the paper's equal-weight 50-configuration ensemble; differences exceeding the stated fitting errors would falsify the central ensemble premise.","tokens_in":13208,"feed_emoji":"🧲","tokens_out":4338,"duration_ms":45389,"temperature":0.7,"pith_summary":"This paper establishes a fully automated protocol for fitting a machine-learned interatomic potential that includes collinear magnetic moments as explicit degrees of freedom. The key ingredients are an active-learning algorithm that selects the most extrapolative atomic configurations on the fly and constrained density functional theory, which supplies energies, forces, stresses, and magnetic forces for those configurations even when the magnetic moments are far from equilibrium. Tested on B1-CrN, the resulting potential reproduces elastic constants, phonon spectra, lattice thermal expansion, and specific heat capacity in the paramagnetic state, matching DFT and experimental values. The authors argue that this removes the manual trial-and-error that has previously dominated the construction of magnetic machine-learning potential training sets.","feed_headline":"Autotrained potential reproduces paramagnetic CrN properties","feed_subtitle":"Elastic constants, phonons, thermal expansion, and heat capacity match DFT and experiment.","key_machinery":"The central object is the magnetic Moment Tensor Potential (mMTP), whose energy is a sum of per-atom contributions built from moment tensor descriptors that depend on interatomic distances, atomic types, and collinear magnetic moments, and whose total energy is explicitly invariant under inversion of all magnetic moments. Fitting uses energies, forces, stresses, and magnetic forces, with active learning based on the D-optimality criterion (maxvol) to select configurations that most increase the linear independence of the fitting matrix. Constrained DFT provides the crucial training labels for non-equilibrium magnetic states by fixing each atom's magnetic moment as an external parameter during self-consistency, which lets the potential learn the energy landscape of excited magnetic configurations.","core_discovery":"The paper's central claim is that an actively trained magnetic Moment Tensor Potential can describe B1-CrN in its paramagnetic state without hand-crafted training sets. The paramagnetic state is represented as an equally weighted average over 50 randomly disordered collinear magnetic configurations, and the fitted potential reproduces the mechanical, dynamical, and thermal properties of this ensemble: elastic constants agree with DFT in the ferromagnetic, antiferromagnetic, and paramagnetic states; phonon spectra agree with DFT and with Raman/infrared measurements once non-analytic LO-TO corrections are included; and quasi-harmonic thermal expansion and heat capacity match experiment. The final training set contains 2423 configurations, roughly 96% of them with non-equilibrium magnetic moments computed with constrained DFT, and prediction errors are 1.7 meV/atom in energy, 108 meV/Å in forces, 0.4 GPa in stresses, and 64 meV/μB in magnetic forces.","pith_inferences":["Extension: the equally weighted 50-configuration ensemble represents the infinite-temperature paramagnetic limit; predictions near the Néel temperature could shift if real finite-temperature paramagnetism carries short-range spin correlations that this ensemble misses.","Extension: the protocol's reliance on constrained DFT means its transferability to other correlated magnets will depend on the choice of constraint spheres and Hubbard U, which would need separate validation for each new material.","Extension: the heat-capacity mismatch near the phase transition suggests that a B1-only training set cannot capture structural phase changes; adding Ortho-CrN data in the same active-learning loop would be a direct test of whether the protocol extends to phase boundaries."],"forward_implications":["The same protocol can build magnetic machine-learning potentials for other materials without manually curated configurations, replacing trial-and-error training-set construction.","One fitted potential simultaneously describes ferromagnetic, antiferromagnetic, and paramagnetic states of B1-CrN, so it can be used for simulations crossing magnetic orderings.","Phonon spectra in the paramagnetic state require non-analytic term corrections to capture LO-TO splitting, and using experimental dielectric and Born charge inputs brings the mMTP spectrum close to Raman and infrared data.","Because anharmonic effects are negligible in CrN, the quasi-harmonic approximation suffices for thermal properties; heat capacity matches experiment above 400 K, while the 280–400 K discrepancy is attributed to the absence of Ortho-CrN data in the training set.","Active learning with constrained DFT is essential in practice: about 96% of the selected training configurations have non-equilibrium magnetic moments that ordinary DFT cannot provide."],"supporting_citations":[{"why":"Supplies the constrained DFT method that fixes magnetic moments as external parameters, enabling training labels for non-equilibrium magnetic configurations.","marker":"[13]"},{"why":"Provides the active-learning algorithm based on D-optimality (maxvol) that the paper generalizes to magnetic moment tensor potentials.","marker":"[12]"},{"why":"Introduces the magnetic Moment Tensor Potential for collinear spin-polarized materials, the functional form extended and automatically trained here.","marker":"[7]"},{"why":"Defines the Moment Tensor Potential class of interatomic potentials on which the magnetic version is built.","marker":"[19]"},{"why":"Shows that fitting to magnetic forces improves magnetic moment tensor potential reliability, justifying the inclusion of magnetic-force targets in the objective function.","marker":"[24]"},{"why":"Establishes the choice of Hubbard U for CrN and the magnetic sampling method for the paramagnetic state, supporting the paper's ensemble averaging approach.","marker":"[17]"},{"why":"Provides DFT elastic constants and phonon spectra for CrN magnetic phases, the quasi-harmonic approximation assessment, and the experimental thermal expansion reference.","marker":"[18]"},{"why":"Supplies experimental phonon frequencies at the Γ point and the dielectric tensor and Born effective charges used for LO-TO corrections in the paramagnetic phonon spectrum.","marker":"[16]"}],"fun_headline_variants":["Active-learning magnetic potential reproduces paramagnetic CrN properties","Autotrained magnetic potential captures CrN's paramagnetic behavior","Self-trained magnetic potential matches CrN paramagnetic state","Active learning builds magnetic potential for paramagnetic CrN","Machine-learned magnetic tensor potential tackles CrN paramagnetism"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paramagnetic state is modeled as an equally weighted average over 50 random collinear magnetic configurations, which is the infinite-temperature limit; if 50 configurations are not representative, or if the actual finite-temperature paramagnetic state contains spin correlations this ensemble cannot capture, the predicted properties could be systematically off.","fun_headline_variants_meta":{"raw":{"variants":["Active-learning magnetic potential reproduces paramagnetic CrN properties","Autotrained magnetic potential captures CrN's paramagnetic behavior","Self-trained magnetic potential matches CrN paramagnetic state","Active learning builds magnetic potential for paramagnetic CrN","Machine-learned magnetic tensor potential tackles CrN paramagnetism"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000773,"raw_usage":{"total_tokens":3377,"prompt_tokens":856,"completion_tokens":2521,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":472,"completion_tokens_details":{"reasoning_tokens":2441}},"tokens_in":472,"tokens_out":2521,"duration_ms":19857,"temperature":1.0,"reasoning_tokens":2441,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:26:29.768258+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the elastic constants, phonon spectra, and thermal expansion of B1-CrN using a paramagnetic ensemble that includes finite-temperature short-range spin correlations (for example, from disordered-local-moment molecular dynamics or Monte Carlo spin flips at realistic temperatures) and compare against the paper's equal-weight 50-configuration ensemble; differences exceeding the stated fitting errors would falsify the central ensemble premise.","supporting_citations":[{"cited_title":"Gonze, B","cited_arxiv_id":null,"evidence_quote":"Supplies the constrained DFT method that fixes magnetic moments as external parameters, enabling training labels for non-equilibrium magnetic configurations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the active-learning algorithm based on D-optimality (maxvol) that the paper generalizes to magnetic moment tensor potentials."},{"cited_title":"Novikov, B","cited_arxiv_id":null,"evidence_quote":"Introduces the magnetic Moment Tensor Potential for collinear spin-polarized materials, the functional form extended and automatically trained here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows that fitting to magnetic forces improves magnetic moment tensor potential reliability, justifying the inclusion of magnetic-force targets in the objective function."},{"cited_title":"Alling, T","cited_arxiv_id":null,"evidence_quote":"Establishes the choice of Hubbard U for CrN and the magnetic sampling method for the paramagnetic state, supporting the paper's ensemble averaging approach."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides DFT elastic constants and phonon spectra for CrN magnetic phases, the quasi-harmonic approximation assessment, and the experimental thermal expansion reference."},{"cited_title":"Zhang and D","cited_arxiv_id":null,"evidence_quote":"Supplies experimental phonon frequencies at the Γ point and the dielectric tensor and Born effective charges used for LO-TO corrections in the paramagnetic phonon spectrum."}],"review_version":1}