{"id":"7df0f4fd-f322-4324-8417-1ed19e1cb926","arxiv_id":"2412.02191","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A neuroevolution machine learning potential for aluminas is trained on DFT data and used to map transition alumina and high-pressure phase boundaries and to search gamma-Al2O3 cation arrangements.","lead":"This paper trains a machine learning interatomic potential for all major forms of aluminum oxide and uses it to compute phase diagrams and explore disputed gamma-alumina structures. It matters because a reliable, fast potential would let researchers simulate alumina under realistic temperatures, pressures, and defect conditions that density functional theory cannot reach.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"High-pressure phase-boundary extrapolation is asserted as proof of comprehensive coverage, but the Rh2O3(II)-CaIrO3 line deviates substantially from reference theory and the abstract's [0,4000] K claim exceeds the demonstrated range.","rationale":"The paper has substantial strengths: a public dataset and code, extensive validation on known phases (EOS, elastic constants, phonons, thermal conductivity, thermal expansion, melt and amorphous structure), and a phase diagram for transitional aluminas that is internally consistent with experimental metastability. The reader's conditional verdict is well aligned with these strengths. The load-bearing weakness is exactly the high-pressure extrapolation: the absence of the high-pressure phases in the training set means the claim of comprehensive coverage is supported only by two phase-boundary comparisons, one of which is quantitatively off by roughly a factor of two in Clapeyron slope and is acknowledged to degrade above 2000 K. This is not an internal inconsistency in the potential construction; it is an overreach in the strength and range of the extrapolation claim. The proposed direct DFT free-energy check would settle whether the remaining discrepancy is a genuine extrapolation failure or merely a known DFT-versus-theory offset. Because the paper's main contributions do not collapse if that check fails, the appropriate disposition remains conditional: the central claims should stand only with either additional validation or a more carefully scoped statement of the extrapolated regime.","tokens_in":18740,"tokens_out":4335,"duration_ms":49622,"concrete_test":"Select two state points on the predicted Rh2O3(II)-to-CaIrO3 boundary, e.g., (2500 K, NEP-predicted pressure) and (3500 K, NEP-predicted pressure), and compute the free-energy difference between the two phases directly with DFT using the same PBEsol settings, via thermodynamic integration or quasi-harmonic approximation. If DFT places the transition more than about 10 GPa away from the NEP line, the abstract's [0,4000] K extrapolation claim should be restricted to the tested regime; if DFT confirms the boundary, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the NEP is a universal potential for aluminas, including the abstract's 'successfully extrapolate the phase diagram ... [0,4000] K and [0,200] GPa', rests on the assumption in Section III C that the active-learning dataset, which contains no Rh2O3(II)- or CaIrO3-type high-pressure phases, has nevertheless sampled the local environments needed for those phases. The only evidence offered is the two computed boundary curves in Fig. 7(b). That evidence is weaker than the claim. The alpha-to-Rh2O3(II) boundary is consistent with experiment and prior theory, but the Rh2O3(II)-to-CaIrO3 boundary deviates substantially from the reference calculation: at 1000 K the reported Clapeyron slope is -4.04 MPa/K versus -9.4 MPa/K from Tsuchiya et al., and the paper itself says accuracy 'diminishes above 2000 K'. The abstract nevertheless advertises success up to 4000 K. Low training RMSE (18.11 meV/atom, 250.88 meV/A) does not certify extrapolation to unseen structures; a potential can fit the training manifold well and fail outside it. Therefore the high-pressure extrapolation is not a settled validation of a 'comprehensively explored' structural space; it is a promising but under-tested extrapolation. If the universality claim is to stand, the high-pressure phases need direct validation rather than inference from one partially matching boundary.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a neuroevolution potential (NEP) for alumina trained on a dataset of 3,335 DFT(PBEsol) configurations covering crystalline polymorphs (α, θ, κ, δ, γ models), amorphous/liquid structures, non-stoichiometric AlxOy, and clusters. Validation includes EOS, elastic constants, phonon dispersions (without LO-TO splitting), thermal conductivity, thermal expansion, melting point, liquid RDF, amorphous quenching, and phase diagrams computed using nonequilibrium thermodynamic integration (NETI). The authors claim the potential enables accurate simulations of various aluminas at larger scales and longer timescales, and that it successfully extrapolates the high-pressure phase diagram of α-Al2O3 to Rh2O3(II)- and CaIrO3-type phases not present in the training set. They also introduce a differential-evolution structure search workflow and apply it to evaluate the Smrčok and Luo models of γ-Al2O3.","tokens_in":19173,"tokens_out":3831,"duration_ms":38636,"significance":"If fully supported, the NEP would be a valuable open resource for the alumina community, and the combination of active learning, NETI-based free-energy calculations, and a structure search workflow is a useful methodological template. The paper reports a broad and mostly careful validation effort (EOS, elastic constants, thermal transport, disordered phases), and the open data and code are clear strengths that increase reproducibility. The main significance hinges on the high-pressure extrapolation claim; the current evidence for that claim is partial, so the paper's impact would be strengthened by direct tests on the high-pressure phases and by aligning the temperature range of the advertised claims with the validated range.","major_comments":[{"comment":"The claim of successful extrapolation over [0, 4000] K is not supported by the evidence presented. The paper itself states that for the Rh2O3(II)-to-CaIrO3 boundary the accuracy 'diminishes above 2000 K,' and at 1000 K the computed Clapeyron slope (-4.04 MPa/K) differs from the reference value (-9.4 MPa/K) by more than a factor of two. Since the abstract advertises a phase diagram up to 4000 K, the authors should either restrict the claim to lower temperatures or provide additional validation at state points above 2000 K, for example direct DFT free-energy calculations at selected thermodynamic conditions along the boundary.","section":"§III.C, Fig. 7(b), and abstract"},{"comment":"The inference that reasonable agreement of one phase boundary with experiment implies that 'the structural space is comprehensively explored' is not a direct test of local-environment coverage. The training set contains no Rh2O3(II)- or CaIrO3-type structures, and a low training RMSE does not certify extrapolation to unseen coordination environments. To support the extrapolation claim, the authors should directly compare NEP and DFT energies, equations of state, or phonons for the high-pressure phases, or provide a descriptor-based distance analysis demonstrating that the active-learning dataset already sampled the relevant local environments.","section":"§III.C, last paragraph"},{"comment":"The computed melting point of 2530 K is roughly 9% above the experimental value of 2327 K. The proposed explanation (defect-free crystal versus real samples with defects, surfaces, and container interfaces) is plausible but not quantitatively supported. Because the paper claims high accuracy at high temperatures, the authors should either provide convergence checks of the two-phase method (e.g., system size, interface orientation, and heating protocol) or temper the claim that the melting point is captured with high accuracy.","section":"§III.B, Fig. 5(a)"}],"minor_comments":[{"comment":"There is a typo: 'we first calculat the melting point' should be 'we first calculate the melting point.'","section":"§III.B, first paragraph"},{"comment":"The sentence 'The inset, which enlarges the region marked by the red dashed lines, is reflected the relative height of the first peak' is grammatically incorrect; 'is reflected' should be 'reflects.'","section":"Fig. 6 caption"},{"comment":"The variant labels jp50s1, jp50s2, jp8bs1, and jp8bs2 are not defined in the main text; a brief explanation or a pointer to the supplementary material would help readers.","section":"§II.A, dataset description"},{"comment":"The reported 2σ ranges for the 16c and 48f Wyckoff positions include negative occupancies (e.g., -0.8–3.3), which are unphysical for site-occupation numbers. If these ranges come from a normal approximation, this should be stated explicitly so that readers do not interpret them as literal allowed occupation counts.","section":"TABLE II"},{"comment":"Reference [71] contains a formatting issue ('Physical review B52' should be 'Physical Review B 52'); several other references also lack consistent journal-name styling.","section":"Reference list"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about high-pressure extrapolation is legitimate and lands on a load-bearing part of the abstract and Section III.C. The paper is otherwise a solid, well-resourced MLIP study. I recommend major revision rather than rejection because the core contribution is defensible; the authors need to add direct validation of the high-pressure phases or explicitly constrain the extrapolation claims to the validated temperature range."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one if you work with alumina or train MLIPs for oxides. It is a serious, mostly well-tested NEP for aluminas covering more phases than any published MLIP I know, and it ships data and code. The core message, that a single MLIP can describe alpha, theta, kappa, delta, gamma, amorphous, and liquid alumina, is credible on the evidence.\n\nWhat is genuinely new: a training set of 3,335 structures including four delta variants and multiple gamma models; phase diagrams for transitional aluminas from NETI free-energy integration, which I do not think has been done for these phases; and a differential-evolution structure search (DELTA) applied to gamma-Al2O3 that reproduces the Smrcok model's cation distribution. The validation suite is unusually broad: EOS, elastic constants, phonons, thermal conductivity, thermal expansion, melting point, liquid RDF, amorphous quenching, and the phase boundaries. The potential's RMSEs are reasonable for such a heterogeneous dataset, and the authors are honest about LO-TO splitting being absent in the phonons.\n\nThe soft spot is the high-pressure extrapolation claim. The abstract advertises successful extrapolation up to 4000 K and 200 GPa. The paper's own numbers do not support that. The computed Rh2O3(II)-CaIrO3 boundary has a Clapeyron slope of -4.04 MPa/K at 1000 K versus -9.4 MPa/K from Tsuchiya et al., and the text says accuracy \"diminishes above 2000 K.\" That is a factor-of-two deviation on a boundary, and the 4000 K claim has no demonstration behind it. The authors also say the structural space is \"comprehensively explored\" on the basis of this one partially matching boundary; that is a stronger conclusion than the evidence warrants. This is an addressable problem: temper the abstract, call the high-pressure part a moderate extrapolation test, and ideally validate one high-pressure phase with direct DFT before publication.\n\nMinor issues: no head-to-head comparison with earlier alumina MLIPs (e.g., Tiwari-Feng for alpha), and the melting point is ~9% high, which the authors attribute to defect-free crystals. Both are minor.\n\nWho gets value: anyone selecting an MLIP for alumina simulations, and people working on phase diagrams of metastable oxides. It deserves a serious referee; the dataset and code availability make it reproducible. I would send it to review, with the expectation that the extrapolation claims get reined in.","headline":"A credible and unusually comprehensive alumina NEP with open data and code; the high-pressure extrapolation and 4000 K claim are oversold and need revision.","tokens_in":19587,"tokens_out":1893,"would_cite":true,"duration_ms":19796,"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 universal machine-learned interatomic potential for aluminas, trained on 3,335 DFT-labelled structures, reproduces known polymorphs and predicts high-pressure phases absent from the training set.","keywords":["machine learning interatomic potential","neuroevolution potential","alumina polymorphs","transitional aluminas","phase diagram","active learning","γ-Al2O3 structure","high-pressure phases"],"falsifier":"A direct test would be to add Rh2O3(II)- and CaIrO3-type configurations to the published training set, retrain, and check whether the extrapolated phase boundaries in Fig. 7(b) move outside the stated uncertainties; if they do, the claim of comprehensive structural-space coverage fails. A shorter check is to compute the free-energy difference between Rh2O3(II)- and CaIrO3-type Al2O3 at high temperature with an independent DFT-based method, since the paper itself reports decreasing accuracy above about 2000 K for that boundary.","tokens_in":18536,"feed_emoji":"💎","tokens_out":7689,"duration_ms":71917,"temperature":0.7,"pith_summary":"Alumina appears in many structures, but the atomic arrangements of its transitional forms remain contested. The paper proposes a single machine-learned interatomic potential, built with the neuroevolution potential (NEP) approach on a dataset of 3,335 DFT-labelled structures selected by active learning, that is accurate across crystalline, amorphous, liquid, and non-stoichiometric aluminas. If the central claim holds, this potential makes near-DFT-quality simulation of aluminas feasible on large systems and long timescales, including the defective transitional phases that experiments cannot fully resolve. It also allows the paper to construct phase diagrams and to evaluate competing structural models of γ-Al2O3, which is the payoff a sympathetic reader would care about.","feed_headline":"One neural potential maps alumina's phases to 4000 K and 200 GPa","feed_subtitle":"Trained on 3,335 DFT structures, the potential reproduces known aluminas and predicts high-pressure phases it never saw.","key_machinery":"The load-bearing object is the neuroevolution potential (NEP), a machine-learned interatomic potential in which the total energy is a sum of atom-centred site energies depending on radial and angular descriptors of the local environment, here built from Chebyshev polynomials and the atomic cluster expansion. Its transferability is manufactured by an active-learning loop that uses farthest point sampling in descriptor space to select the most representative new configurations from molecular dynamics runs, labels them with DFT, and iterates until the potential stops improving. Free-energy phase boundaries are then computed with nonequilibrium thermodynamic integration, and the γ-Al2O3 model assessment is carried by a differential-evolution structure search that varies cation occupancies over the allowed Wyckoff sites and uses NEP energies as the stability criterion.","core_discovery":"On its own terms, the paper's discovery is that the NEP captures the energy landscape of alumina well enough to be a universal potential: energy, force, and stress root-mean-square errors of 18.11 meV/atom, 250.88 meV/Å, and 81.01 MPa on the training set, with agreement to DFT equations of state, elastic constants, phonons (except LO-TO splitting from the lack of non-analytical corrections), thermal expansion, thermal conductivity, and the radial distribution functions of liquid and amorphous alumina. In the phase-diagram part, the NEP yields anharmonic-free-energy phase boundaries among α-, θ-, δ-, κ-, and γ-Al2O3 and, despite containing no high-pressure phases, correctly predicts negative Clapeyron slopes for the α→Rh2O3(II) and Rh2O3(II)→CaIrO3 transitions, with the former boundary agreeing with experiments near 90 GPa and the latter becoming less accurate above 2000 K. In the structure-search part, the potential favors a cation distribution with a roughly 97:3 ratio of spinel to non-spinel sites under the Smrčok model, slightly above the 94:6 ratio in the original model, and disfavors the nearest-neighbour cation pairs present in the Luo model.","pith_inferences":["The success of extrapolation here implies a general principle for building machine-learned potentials: explicit coverage of local atomic environments, rather than the list of phases in the training set, determines transferability to unseen structures.","Because the NEP lacks non-analytical corrections for long-range Coulomb interactions, its vibrational and thermal properties are most trustworthy where polarization effects are not decisive; the paper's thermal-conductivity agreement suggests this limitation is minor for α-Al2O3, but an extension with explicit charge terms could close the gap in ionic phonon dispersions.","If the 97:3 spinel-to-non-spinel ratio is robust, it provides an energetically anchored prediction for γ-Al2O3 that could be tested by revisiting electron-diffraction refinements with the NEP-optimized occupancy distribution, since experimental samples may retain local constraints that prevent full relaxation.","The same active-learning-plus-search workflow could be applied to other vacancy-structured oxides, such as spinel ferrites or defective catalytically active oxides, where cation-distribution combinatorics block direct DFT enumeration."],"forward_implications":["Alumina simulations that previously required empirical potentials can now be done at near-DFT accuracy over nanosecond timescales and tens of thousands of atoms, covering crystalline, amorphous, liquid, and non-stoichiometric AlxOy.","The NEP phase diagram gives concrete, anharmonic free-energy-based phase boundaries among α-, θ-, δ-, and κ-Al2O3, including near-parallel δ/θ boundaries that explain observed phase coexistence.","The moderate extrapolation to 4000 K and 200 GPa predicts α→Rh2O3(II) and Rh2O3(II)→CaIrO3 boundaries with correct negative Clapeyron slopes, supporting use of the potential outside the training regime.","The differential-evolution workflow supplies a NEP-based criterion for judging γ-Al2O3 models: spinel-site occupancy ratio of about 97:3 is energetically favored under the Smrčok model, and low-energy structures avoid nearest-neighbor cation pairs.","The potential, dataset, and structure search code are released, enabling other groups to extend coverage to χ- and η-alumina and to refine high-pressure structures."],"supporting_citations":[{"why":"Defines the neuroevolution potential method that the paper trains.","marker":"[42]"},{"why":"Provides the NEP implementation and descriptors used to fit and run the potential.","marker":"[44]"},{"why":"Supplies the energy-locality hypothesis that justifies the atom-centred NEP site-energy ansatz.","marker":"[16]"},{"why":"Introduces the farthest point sampling scheme used to prune active-learning configurations.","marker":"[65]"},{"why":"Establishes the nonequilibrium thermodynamic integration approach used to compute phase boundaries.","marker":"[45]"},{"why":"Defines the Smrčok γ-Al2O3 model whose cation distribution the structure search evaluates.","marker":"[48]"},{"why":"Defines the Luo γ-Al2O3 model whose energetic plausibility is tested against the search results.","marker":"[49]"},{"why":"Provides the DFT-based high-pressure phase boundaries and Clapeyron slopes used as a benchmark for the extrapolation.","marker":"[97]"},{"why":"Supplies experimental phase-boundary data for corundum, Rh2O3(II), and CaIrO3-type Al2O3 in Fig. 7(b).","marker":"[100]"},{"why":"Supplies the differential evolution algorithm underlying the γ-Al2O3 structure search workflow.","marker":"[110]"}],"fun_headline_variants":["Neural potential maps alumina phases to 200 GPa","Machine learning charts alumina's full phase diagram","One MLIP predicts alumina's extreme-condition phases","Alumina's energy landscape decoded by a neural potential","ML model reproduces and predicts alumina structures"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the training set, which contains no high-pressure Rh2O3(II)- or CaIrO3-type alumina, nevertheless covers the local atomic environments needed for the potential to make trustworthy predictions about those phases; if a high-pressure coordination environment was under-represented, the extrapolated phase boundaries would be unreliable rather than revealing.","fun_headline_variants_meta":{"raw":{"variants":["Neural potential maps alumina phases to 200 GPa","Machine learning charts alumina's full phase diagram","One MLIP predicts alumina's extreme-condition phases","Alumina's energy landscape decoded by a neural potential","ML model reproduces and predicts alumina structures"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000336,"raw_usage":{"total_tokens":1933,"prompt_tokens":1088,"completion_tokens":845,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":704,"completion_tokens_details":{"reasoning_tokens":773}},"tokens_in":704,"tokens_out":845,"duration_ms":8869,"temperature":1.0,"reasoning_tokens":773,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:44:11.943739+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would be to add Rh2O3(II)- and CaIrO3-type configurations to the published training set, retrain, and check whether the extrapolated phase boundaries in Fig. 7(b) move outside the stated uncertainties; if they do, the claim of comprehensive structural-space coverage fails. A shorter check is to compute the free-energy difference between Rh2O3(II)- and CaIrO3-type Al2O3 at high temperature with an independent DFT-based method, since the paper itself reports decreasing accuracy above about 2000 K for that boundary.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the neuroevolution potential method that the paper trains."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the NEP implementation and descriptors used to fit and run the potential."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the farthest point sampling scheme used to prune active-learning configurations."},{"cited_title":"de Koning, A","cited_arxiv_id":null,"evidence_quote":"Establishes the nonequilibrium thermodynamic integration approach used to compute phase boundaries."},{"cited_title":"Smrˇ cok, V","cited_arxiv_id":null,"evidence_quote":"Defines the Smrčok γ-Al2O3 model whose cation distribution the structure search evaluates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Luo γ-Al2O3 model whose energetic plausibility is tested against the search results."},{"cited_title":"Tsuchiya, T","cited_arxiv_id":null,"evidence_quote":"Provides the DFT-based high-pressure phase boundaries and Clapeyron slopes used as a benchmark for the extrapolation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies experimental phase-boundary data for corundum, Rh2O3(II), and CaIrO3-type Al2O3 in Fig. 7(b)."}],"review_version":1}