{"id":"4e2948ef-9f5c-4d91-bf14-878c8982a126","arxiv_id":"2607.08351","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"Three foundation MLIPs succeed on scalar breathing order, fail the symmetry of vector Jahn–Teller order, and cannot access on-site multiplet crossovers in d4/d6/d7 perovskites.","lead":"Foundation ML potentials capture breathing-mode order in NdNiO3, get Jahn–Teller magnitude but wrong symmetry in LaMnO3, and cannot see the LaCoO3 spin crossover. The work ranks which correlated lattice orders today’s structure-only models can and cannot learn.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged soft spots on the 'iff' criterion and LCO inaccessibility.","rationale":"The reader's CONDITIONAL verdict already isolates the precise soft spots: data-on-request, the strong 'iff' strained by omol-NNO, and LCO inaccessibility argued from absence of geometric OP rather than a failed multiplet baseline. After re-reading the full text (protocol, Methods descriptors, Tables 1-2, Figs. 1-2, finite-size checks, hierarchy), I find no additional load-bearing concern that would move the verdict. The scalar/vector/on-site hierarchy is cleanly supported by the reported Q_RS, |Q_static_2|, BC, M_FO/M_OO values across three potentials and two cell sizes. The geometric definition of 'orbital order' is explicitly flagged as a structural surrogate. Softening the criterion wording and releasing trajectories would convert CONDITIONAL to ACCEPT; no internal inconsistency or untested assumption stronger than those already noted exists. Hence UNCHANGED and full agreement with the reader.","tokens_in":12514,"tokens_out":653,"duration_ms":6354,"concrete_test":"Release the full T-series trajectories and recompute Table 2 / Fig. 2 classifiers after adding one multiplet-aware fine-tune (or explicit local-moment target) on LaCoO3 only; if the LCO bond-mean PDF and thermal expansion remain featureless while NNO/LMO classifiers are unchanged, the 'fundamentally inaccessible' claim is strengthened; if a clear LS/HS bimodality or expansion appears, soften the on-site class language.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (Structural-footprint criterion and Hierarchy sections) is that foundation MLIPs reproduce static low-T fingerprints iff the electronic instability has condensed onto a training-set lattice distortion, mapping NNO (scalar, captured), LMO (vector, magnitude not symmetry), and LCO (on-site, inaccessible). The reader's weakest_assumption correctly flags the two softest points: (i) LCO multiplet crossover is declared fundamentally unreachable from any structure-only model solely because it lacks a spatial order parameter (rather than being tested against a multiplet-aware baseline), and (ii) geometric surrogates Q_RS and long-axis M_OO/M_FO are taken as full diagnostics of the electronic channels. These are real but already identified; they do not create an internal contradiction. The multi-potential, multi-T, 80/160-atom MD results consistently support the differential hierarchy that is actually demonstrated. The 'iff' wording is slightly strained by omol freezing NNO breathing without solid-state training, but the paper itself notes this is a kinetic basin, not a thermodynamic T_MI prediction. No stronger load-bearing flaw (e.g., finite-size aliasing, descriptor circularity, or misclassification of FO vs C-type) is present in the reported data.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript tests three foundation machine-learning interatomic potentials (MACE matpes_r2scan, MACE omol, and CHGNet) on the low-temperature structural order of three isostructural ABO3 perovskites—LaMnO3 (d4), LaCoO3 (d6), and NdNiO3 (d7)—via identical NVT molecular-dynamics protocols (1 ns production, 50–300 K, 80- and 160-atom supercells) with no system-specific training. It reports a hierarchy of difficulty: NdNiO3’s scalar rocksalt breathing fingerprint is captured (fully by omol, as a precursor by matpes_r2scan); LaMnO3’s vector Jahn–Teller long-axis order is captured in magnitude by some potentials but not in the experimental C-type symmetry (ferro-orbital patterns appear instead); and LaCoO3’s low-spin to high-spin multiplet crossover, which has no spatial order parameter, leaves no static structural footprint and is not recovered. From these outcomes the authors propose a structural-footprint criterion: foundation MLIPs trained on charge- and spin-converged DFT total energies reproduce static low-T fingerprints if and only if the electronic instability has condensed onto a lattice distortion already present in the training-set relaxations.","tokens_in":12859,"tokens_out":1782,"duration_ms":27216,"significance":"The work is timely: foundation MLIPs are being adopted for finite-temperature structure and dynamics in correlated oxides, yet their fidelity for multiplet, orbital, and bond-disproportionation physics is poorly mapped. The comparative design—three materials chosen to isolate scalar, vector, and on-site channels; three potentials spanning closed-shell, r2SCAN, and GGA+U/magnetism-supervised training; matched 1 ns trajectories; and explicit 160-atom finite-size checks that leave every classifier on the same side of the reported thresholds—is a genuine strength and makes the differential hierarchy credible. The paper also states falsifiable, prospective predictions for the wider RNiO3, Co multiplet-crossover, and cooperative-JT families, and documents a reusable descriptor pipeline and trajectory availability. If the hierarchy and the more carefully worded form of the structural-footprint criterion hold, they give the community a concrete, material-specific checklist for when structure-only foundation potentials can be trusted and when multiplet- or spin-aware fine-tuning is mandatory.","major_comments":[{"comment":"Structural-footprint criterion (main text, paragraph beginning “The cross-MLIP outcome … maps cleanly onto a single principle”): the “if and only if” wording is stronger than the evidence. The three potentials and three materials support a useful one-way implication (condensed training-set distortion → capture of the static fingerprint is possible), but the converse and the universality claim rest on a small sample. In particular, omol freezes the NdNiO3 rocksalt pattern (Q_RS = −1.000) despite no extended-solid cooperative-distortion training; the paper correctly notes this is a kinetic basin rather than a thermodynamic T_MI prediction, yet that success without the stated training-set content already strains the “already present in the training-set ground-state relaxations” half of the criterion. Soften to a working hypothesis or one-directional rule, and state the domain of tested pote","section":"Structural-footprint criterion"},{"comment":"Hierarchy of broken-symmetry channels and Outlook: the claim that the LaCoO3 multiplet crossover is “fundamentally unreachable from any structure-only model” and that “no enrichment of bond-based geometric features can recover it” goes beyond what is demonstrated. The MD results show that the three present-day foundation potentials return no static order-parameter-locked distortion and only mild leptokurtic or ringing signatures, which is consistent with experimental absence of a spatial OP and with the abstract’s “inaccessible to present-day MLIPs.” That is sufficient and important. Declaring architectural impossibility for all future structure-only models, without a multiplet-aware or multi-reference baseline comparison, is not load-bearing for the hierarchy actually shown and should be rephrased as a strong expectation requiring multiplet-resolved auxiliary inputs (as the Outlook alre","section":"Hierarchy of broken-symmetry channels / Outlook"},{"comment":"Methods (“orbital order” paragraph) and LaMnO3 results: throughout, M_OO / M_FO and “orbital order” are structural surrogates (long-axis C-type vs ferro-orbital patterns) with no electronic-orbital information in the potentials. This is stated once, but the abstract, Table 2, and hierarchy language still read as if the electronic e_g channel itself has been diagnosed. For the central claim this is acceptable only if every success/failure statement is explicitly about the structural fingerprint (as Table 1’s “static fingerprint” column does). Please audit the abstract and the LMO rows of Table 2 so that “captured in magnitude but not in symmetry” is unambiguously the geometric long-axis pattern, not a claim about orbital occupation.","section":"Methods; Abstract; Table 2"}],"minor_comments":[{"comment":"Figure 1 caption and panel (c): shell correlations are shown only at 50 K; a brief note that the (- + - +) alternation for NdNiO3 persists across the full T series (as stated in the text) would help readers who look only at the figure.","section":"Figure 1"},{"comment":"Table 2 caption: “≈expt” is carefully caveated for the structural channel only; consider adding the same caveat in the LaCoO3 matpes_r2scan cell text so that “≈expt: Gaussian, no distortion” is not misread as electronic agreement with the spin crossover.","section":"Table 2"},{"comment":"Eq. (1) and the long-axis mask |Q_static_2| < 0.05 Å: both are free parameters of the analysis. A one-sentence sensitivity check (e.g. that FO vs C-type classification is stable for a small window around 0.05 Å) would strengthen Methods; the 160-atom check already helps on finite size.","section":"Methods"},{"comment":"Notation consistency: matpes_r2scan / matpes r2scan / r2scan appear with and without underscores; pick one form for the production head throughout text and figures.","section":"Throughout"},{"comment":"Data availability: “available from the corresponding author on reasonable request” is weaker than the Outlook’s claim that trajectories and the descriptor pipeline are made available to enable comparison. If a repository deposit is planned, state it; if not, align the Outlook wording with the Methods statement.","section":"Data and code availability / Outlook"}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is a solid, well-executed comparative study and is a good fit for a materials/condensed-matter journal that values methodological stress-tests of emerging ML tools. The soft spots are interpretive overreach (iff criterion; “fundamentally unreachable”), not data or protocol failures; minor revision with language tightening should be enough. I would not require new multiplet-aware MD for acceptance of the hierarchy that is actually demonstrated."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful part of this paper is the controlled comparison: three foundation potentials, three isostructural d4/d6/d7 perovskites, identical 1 ns NVT protocols on 80- and 160-atom cells, no system-specific training, and a small set of geometric classifiers (Q_RS, |Q2|, BC, long-axis FO vs C-type). That is new. Prior work gave us MACE and CHGNet and lots of DFT/AIMD on these oxides; it did not give us this side-by-side hierarchy.\n\nWhat they show is clear and reproducible from the reported numbers. NdNiO3 breathing (scalar) is captured, fully by omol and as a precursor by the others. LaMnO3 JT magnitude can appear (especially CHGNet) but the long-axis pattern is ferro-orbital, not experimental C-type. LaCoO3 shows no symmetry-breaking footprint, which matches the experimental absence of a static order parameter. Finite-size checks keep every classifier on the same side of the thresholds. Descriptors are defined, applied consistently, and the FO/C-type distinction is not ambiguous in the data they report. The outlook is concrete: what training data or auxiliary labels would be needed for each class.\n\nSoft spots are real but limited. The strong “iff” footprint criterion is a bit over-tightened by omol freezing the NNO rocksalt pattern without solid-state training (they themselves call it a kinetic basin). Declaring the LCO multiplet channel “fundamentally unreachable” from any structure-only model rests on the absence of a spatial order parameter rather than a failed multiplet-aware baseline; that is an argument, not a proof. Primary trajectories are on-request only. None of these break the comparative hierarchy that is actually demonstrated.\n\nThis is for people who run or train foundation potentials on correlated oxides and need a practical filter, not a new potential. Math and citation pattern look solid; self-cites supply context. I would send it to referees. Soften the “iff” and LCO wording, release the trajectories, and it is a useful short paper.","headline":"Clean ns-scale MD benchmark that turns three perovskites into a practical scalar/vector/on-site filter for when foundation MLIPs can stand in for DFT dynamics.","tokens_in":13497,"tokens_out":530,"would_cite":true,"duration_ms":5385,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Foundation machine-learning potentials capture perovskite order only when it freezes into geometry already in their training data, ordered by a scalar-vector-on-site hierarchy of difficulty.","keywords":["foundation machine-learning interatomic potentials","perovskite oxides","Jahn-Teller distortion","breathing mode","spin-state crossover","orbital order","molecular dynamics","strongly correlated electrons"],"falsifier":"Run the same nanosecond molecular-dynamics protocol and descriptor pipeline on additional foundation potentials or on fine-tuned models that include explicit multiplet or local-spin labels for LaCoO3 (or correct C-type Jahn-Teller patterns for LaMnO3) and check whether the reported order parameters, bimodality coefficients and long-axis classifications reverse.","tokens_in":13383,"feed_emoji":"⚛️","tokens_out":769,"duration_ms":7087,"temperature":0.7,"pith_summary":"Foundation machine-learning interatomic potentials are replacing density-functional theory for structure and nuclear dynamics, yet their reliability for strongly correlated oxides remains open. This paper tests three such potentials, with no material-specific training, on nanosecond molecular dynamics of three isostructural perovskites that host textbook-different low-temperature orders: NdNiO3 (breathing-mode bond disproportionation), LaMnO3 (cooperative Jahn-Teller orbital order), and LaCoO3 (site-local spin multiplet crossover). The materials map onto three classes of increasing architectural demand. A scalar collective bond mode is captured; a vector field of long-bond axes is captured in magnitude but lands on the wrong symmetry pattern; a purely on-site multiplet population shift that leaves no static geometric order parameter is invisible. The paper therefore supplies a structural-footprint criterion: these potentials reproduce static low-temperature fingerprints if and only if the electronic instability has already condensed its order parameter onto a lattice distortion present in the training-set ground states. The hierarchy points to concrete, material-specific training upgrades that can be checked against the same diagnostics.","feed_headline":"ML potentials capture oxide order only when geometry already freezes it","feed_subtitle":"Scalar bond modes work; vector orbital patterns get the wrong symmetry; on-site spin crossovers stay invisible.","key_machinery":"The structural-footprint criterion together with the scalar-vector-on-site hierarchy: scalar bond-mode order is learnable from nearest-neighbour anti-correlations alone; vector long-axis order requires the correct symmetry pattern in the training set; on-site multiplet population has no geometric order parameter and cannot be recovered from bond-based features.","core_discovery":"A foundation MLIP trained only on charge- and spin-converged DFT total energies, forces and stresses will reliably reproduce the static low-temperature structural fingerprints of a correlated oxide if and only if the underlying electronic instability has condensed its order parameter onto a static lattice distortion already present in the training-set ground-state relaxations. The three perovskites instantiate a hierarchy of difficulty: scalar (NdNiO3 rocksalt breathing, captured), vector (LaMnO3 long-axis orbital order, magnitude sometimes correct but symmetry wrong), and on-site (LaCoO3 multiplet crossover, inaccessible).","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["MLIPs capture oxide order only if lattice already freezes the electronic instability","Scalar bond modes work for foundation MLIPs; vector symmetry and spins fail","Foundation MLIPs reproduce order solely when it condenses onto static lattice distortions","NdNiO3 breathing captured; LaMnO3 orbital symmetry wrong; LaCoO3 spins invisible","Hierarchy for MLIPs: scalar yes, vector magnitude only, on-site multiplet crossover inacce"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The claim that the LaCoO3 spin crossover is fundamentally unreachable from any structure-only model, rather than merely missed by today's geometric descriptors, and that purely structural surrogates fully diagnose success or failure for the electronic channels.","fun_headline_variants_meta":{"raw":{"variants":["MLIPs capture oxide order only if lattice already freezes the electronic instability","Scalar bond modes work for foundation MLIPs; vector symmetry and spins fail","Foundation MLIPs reproduce order solely when it condenses onto static lattice distortions","NdNiO3 breathing captured; LaMnO3 orbital symmetry wrong; LaCoO3 spins invisible","Hierarchy for MLIPs: scalar yes, vector magnitude only, on-site multiplet crossover inaccessible"]},"model":"grok-4.5","effort":"low","cost_usd":0.00711,"raw_usage":{"total_tokens":1801,"prompt_tokens":828,"num_sources_used":0,"completion_tokens":114,"cost_in_usd_ticks":71100000,"prompt_tokens_details":{"text_tokens":828,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":859,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":828,"tokens_out":114,"duration_ms":7908,"temperature":1.0,"reasoning_tokens":859,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T08:52:44.763776+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the same nanosecond molecular-dynamics protocol and descriptor pipeline on additional foundation potentials or on fine-tuned models that include explicit multiplet or local-spin labels for LaCoO3 (or correct C-type Jahn-Teller patterns for LaMnO3) and check whether the reported order parameters, bimodality coefficients and long-axis classifications reverse.","supporting_citations":[],"review_version":1}