Pith. sign in

REVIEW 3 major objections 5 minor 6 references

PASS: Perturbation augmented space group structure sampling for transferable Fe-O machine learning interatomic potential

T0 review · 3 major / 5 minor · reviewed 2026-07-31 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Small-cell space-group sampling plus controlled perturbations trains a transferable Fe–O potential that grows FeO-like oxide in large simulations.

desk verdict Useful ASSYST extension plus a real Fe–O ACE validation suite; the oxidation “FeO-like” read is structural, not a clean phase proof under fixed-FM thermodynamics. read the letter →

arxiv 2607.28000 v1 pith:EFGCMYFU submitted 2026-07-30 cond-mat.mtrl-sci physics.comp-ph

Zixiong Wei , Fei Shuang , Poulumi Dey This is my paper · ORCID
classification cond-mat.mtrl-sciphysics.comp-ph
keywords IronoxidationSpacegroupsMachinelearninginteratomicpotentialsAtomicclusterexpansionMoleculardynamicsPASSsamplingFe-Osystem
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Iron oxidation is hard to model atomistically because the Fe–O system mixes many crystal structures, stoichiometries, defects, and magnetic orderings, so building a trustworthy training set for a machine-learning potential has been slow and incomplete. This paper introduces PASS: generate compact crystals from space groups (cells under 10 atoms), relax them, then systematically rattle atoms and stretch cells to create non-equilibrium local environments, and keep only the most representative ones via SOAP descriptors and farthest-point sampling. An atomic-cluster-expansion potential trained on that one-shot set is then shown to match bulk, surface, defect, adsorption, diffusion, and interface energetics for pure iron and common oxides—even though those large or defective configurations were never put in the training data. In a nanosecond-scale molecular-dynamics run of oxygen on BCC iron at 873 K, the same potential forms a surface layer whose composition and short-range order are FeO-like. The practical claim is that carefully sampled small cells can transfer to chemically complex oxide growth without iterative active learning or hand-built oxidation pathways.

What carries the argument

PASS (Perturbation Augmented Space group structure Sampling): space-group prototypes are volume- and fully relaxed, then extensively rattled and stretched on a grid of strain parameters; SOAP descriptors plus PCA and farthest-point sampling retain a diverse, non-redundant subset of local atomic environments for one-shot DFT labelling and ACE training.

What would settle it

Run the same large-scale BCC Fe oxidation MD with a spin-aware or DFT+U-referenced potential (or compare directly to in-situ structural probes) and check whether the surface layer still reaches near-equal Fe/O composition with FeO-like RDF peak positions and coordination numbers; systematic mismatch would falsify the claimed transferability.

Watch

Extended reading notes

Core claim

PASS produces a compact, representative Fe–O training set of small-cell structures such that an ACE machine-learning interatomic potential trained on it is accurate and transferable enough to reproduce bulk, surface, and interface thermodynamics and kinetics and to capture the reactive complexity of large-scale Fe oxidation, including formation of an FeO-like oxide layer.

Load-bearing premise

Local environments taken from heavily perturbed cells of at most ten atoms, labelled under one fixed ferromagnetic DFT setting and without explicit magnetism or long-range electrostatics, are assumed to be enough for the energetics and kinetics of real iron oxides, interfaces, and high-temperature oxidation.

Editorial extensions

If this is right

  • A transferable Fe–O ACE potential can be built from small cells alone, cutting the cost of high-fidelity DFT labelling relative to large hand-crafted or actively learned sets.
  • Large-scale MD of early oxide growth becomes practical and can spontaneously form FeO-like short-range order under high-temperature conditions.
  • The same sampling logic can be reused for other chemically complex reactive systems where manual pathway enumeration or many active-learning cycles are currently required.
  • Defect, adsorption, diffusion, and interface properties of Fe and Fe oxides can be screened without putting those configurations into the original training set.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because magnetism is fixed and long-range interactions are cut off, regimes where AFM/ferrimagnetic order or spin disorder dominate (Curie/Néel crossings, certain vacancy barriers) remain the natural next stress tests.
  • PASS-style one-shot sampling could be paired later with targeted active learning only for rare defective interfaces rather than for the whole configurational space.
  • If the FeO-like layer result holds under broader oxygen pressures and temperatures, the potential becomes a practical tool for corrosion and iron-powder energy-storage modelling.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The manuscript introduces PASS, a one-shot workflow that generates compact Fe and Fe–O training sets from space-group prototypes (≤10 atoms), stepwise relaxation, controlled rattle/stretch perturbations, SOAP+PCA embedding, and farthest-point sampling, then trains an ACE MLIP under a fixed ferromagnetic DFT labeling protocol. The potential is validated on pure-Fe EOS, thermal expansion, Bain path, cleavage/T–S and GSF energetics, point defects and surfaces, and on Fe–O bulk oxides, surfaces, Fe/FeO interfaces, O adsorption/dissolution, O diffusivity versus experiment, and vacancy formation/migration barriers, before a 1 ns BCC Fe(100) oxidation MD at 873 K that produces a near-equiatomic surface layer whose RDF and first-shell coordination are interpreted as FeO-like. The central claim is that small-cell LAE sampling alone yields a transferable Fe–O MLIP capable of reactive oxide-growth complexity without iterative active learning or explicit large oxidation structures in training.

Significance. If the transferability claim holds, PASS would be a practical, symmetry-informed alternative to manual enumeration, RSS, and multi-cycle active learning for chemically complex reactive systems, and a first demonstration that small-cell space-group sampling can support large-scale Fe oxidation MD. Strengths include the systematic out-of-distribution validation suite (defects, GSF, adsorption, CINEB, experimental O diffusion, interfaces) and the explicit comparison of SOAP embeddings against application-specific Fe and Fe–O datasets. The work is of clear interest for MLIP dataset design and high-temperature oxidation modeling, provided thermodynamic biases and the magnetic limitation are quantified rather than only acknowledged.

major comments (3)
  1. [Table 3; Fig. 4(d)] Table 3 and Fig. 4(d): formation enthalpies of FeO, Fe2O3 and Fe3O4 are systematically more exothermic than DFT(+U) references (e.g. FeO −1.57 vs −0.91/−1.43 eV/atom), and Fe vacancy formation energies in FeO/Fe3O4 are underestimated even though trends are preserved. These are load-bearing for the claim that the potential captures oxide thermodynamics. The manuscript should quantify how these biases affect relative oxide stability (with and without U, and versus O chemical potential) and state clearly which thermodynamic conclusions remain reliable.
  2. [Fig. 5; Application to large-scale oxidation simulation] Fig. 5 oxidation showcase: the “FeO-like” identification rests on ~50% composition, broadened RDF peak positions, and first-shell CN from a single 873 K trajectory, without energy ranking against Fe2O3/Fe3O4-like order, without magnetic order, and without a control that the biased ΔHf does not preferentially stabilize rock-salt short-range order. Given the Discussion’s own limits on magnetism and the Table 3 offsets, either add comparative structural/thermodynamic diagnostics or soften the claim from phase-like identification to short-range Fe–O order under the simulated conditions.
  3. [Methods (DFT calculations); Discussion; Fig. 3(c)] Methods and Discussion: all training labels use fixed ferromagnetic ordering without U or explicit spin degrees of freedom, while target oxides are AFM/ferrimagnetic and high-T oxidation involves spin disorder. Bain-path (Fig. 3c) and oxide tests only partially probe this gap. A concrete statement of the domain of applicability—and, if feasible, a limited spin-ordered or +U comparison on key oxide/interface energies—is needed so readers can judge when the structural-transferability argument is sufficient.
minor comments (5)
  1. [Systematic validation of the ACE MLIP for pure Fe] Training errors (31.89 meV/atom, 120.52 meV/Å) are given without a held-out test split or learning-curve context; a brief test-set or bootstrap estimate would help interpret absolute accuracy.
  2. [The PASS workflow] Supplementary Table 1 (perturbation grid) is cited as central to PASS reproducibility but is not in the main text; ensure the grid and FPS size are fully specified for reuse.
  3. [Fig. 2] Fig. 2(c,d) PCA overlays are illustrative; state explicitly that coverage is necessary but not sufficient evidence of transferability (the property tests carry that burden).
  4. [References] Minor typographical issues appear in the reference list and elsewhere (e.g. oxida3on-style character substitutions, “V olmin”, “plaroorm”); a full proofread is needed.
  5. [Discussion; Methods] ACE cutoff (6 Å) and max cell size (<10 atoms) imply truncated long-range electrostatics in ionic oxides; a short note on implications for charged defects and polar surfaces would help.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: PASS→ACE is standard supervised fitting; transferability is checked on external DFT/experiment, not forced by construction.

full rationale

The load-bearing chain is: (1) generate small-cell space-group + rattle/stretch structures; (2) SOAP/PCA/FPS down-select; (3) label with fixed-FM DFT; (4) fit ACE; (5) validate on bulk/surface/defect/oxide/interface/adsorption/CINEB/O-diffusion and a large oxidation MD. None of these steps defines the claimed outputs in terms of the fitted inputs. Oxide lattices, formation enthalpies, vacancy barriers, O adsorption/dissolution energies, experimental O diffusivity, and the FeO-like MD morphology are not parameters of the PASS selection or the ACE loss; several are explicitly stated as absent from training. PCA coverage plots (Fig. 2) are post-hoc illustrations against other datasets, not the selection objective that forces later property matches. Method inheritance from ASSYST (Poul et al.) and reuse of the authors’ prior DFT adsorption set as a test benchmark are ordinary practice and do not make the central transferability claim true by definition. Thermodynamic biases vs DFT(+U) and the non-spin-aware limitation are accuracy/scope issues, not circular reductions. Score 1 only for mild same-stack DFT train/test dependence typical of MLIP papers.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

The load-bearing story is methodological and empirical: space-group + perturbation coverage of LAEs plus ACE regression under standard DFT labels. Free choices are the PASS mesh, SOAP/FPS selection sizes, ACE architecture/cutoff, and the fixed-FM labeling policy. No new physical entities are postulated; domain assumptions are standard DFT/MD practice plus the transferability hypothesis that small-cell LAEs dominate oxide-growth physics.

free parameters (6)
  • Rattle/stretch perturbation grid (σ_rattle, ε_hydro, ε_shear, n_rattle, n_stretch) = Grid in Supp. Table 1 (not fully numeric in main text)
    Hand-chosen mesh that defines the non-equilibrium LAE coverage; values deferred to Supplementary Table 1 and directly control what FPS can select.
  • FPS selection size = 10000 per system
    10,000 perturbed structures each for Fe and Fe–O after SOAP–PCA; arbitrary compactness knob.
  • Max atoms per space-group cell = 10
    Hard cap of 10 atoms sets the small-cell regime and neglect of longer-range motifs.
  • SOAP hyperparameters (r_cut, n_max, l_max, σ) = r_cut=7 Å, n_max=12, l_max=4, σ=0.3
    Descriptor used for selection (not the ACE fit itself); choices change which structures are kept.
  • ACE model hyperparameters = cutoff 6 Å; 800 B-functions/element; weight 1e-2
    800 B-basis functions/element, 4448 parameters, 6 Å cutoff, force/energy loss weight 1e-2, L1/L2=1e-8, 4000 BFGS steps—standard fit knobs that set capacity and bias.
  • Minimum interatomic distance filters = 1.8 Å Fe; 1.6 Å Fe–O
    1.8 Å (Fe) and 1.6 Å (Fe–O) cutoffs discard structures before labeling; affects force distribution and physicality.
assumptions (5)
  • domain assumption Collinear spin-polarized PBE-PAW DFT (and selective +U only in literature comparisons) is an adequate label of Fe–O energetics for MLIP training and validation.
    Methods DFT section; oxide tables mix DFT and DFT(+U) references while training appears without systematic +U.
  • ad hoc to paper A single ferromagnetic ordering for Fe in all training labels is sufficient for a structurally transferable potential even when target oxides are AFM or ferrimagnetic.
    Stated explicitly before pure-Fe validation and revisited in Discussion as a limitation.
  • domain assumption Local atomic environments encoded within a finite cutoff (SOAP 7 Å selection; ACE 6 Å) dominate the PES relevant to oxidation, so small cells transfer to surfaces, interfaces, and MD growth.
    Core PASS transferability hypothesis in Results/Discussion; long-range interactions acknowledged as neglected.
  • domain assumption ACE with Finnis–Sinclair-type embedding and chosen B-basis is expressive enough to regress the labeled PES without explicit bond-order or charge equilibration terms.
    MLIP training Methods; reactive complexity is claimed from geometry-based ACE alone.
  • domain assumption Space-group enumeration via PyXtal plus ASSYST-style rattle/stretch yields a controlled cover of crystalline and near-crystalline LAEs relevant to defects and reaction paths.
    PASS workflow builds directly on Poul et al. ASSYST citations.
invented entities (1)
  • PASS (Perturbation Augmented Space group structure Sampling) workflow
    purpose: Name the combined one-shot pipeline: space-group generation, stepwise relax, perturbation mesh, SOAP–PCA–FPS selection for MLIP datasets.
    Methodological construct, not a physical object; independent use would be other materials’ training sets built the same way.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PASS: Perturbation augmented space group structure sampling for transferable Fe-O machine learning interatomic potential." pith.science (2026). https://pith.science/paper/EFGCMYFU

@misc{pith2026260728000,
  author       = {Pith},
  title        = {Pith review of: PASS: Perturbation augmented space group structure sampling for transferable Fe-O machine learning interatomic potential},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EFGCMYFU}},
  note         = {Machine review of arXiv:2607.28000}
}
read the original abstract

Accurate atomistic modelling of iron (Fe) oxidation requires a reliable interatomic potential, which necessitates an extensive and representative first-principles dataset for training the interatomic potential. However, Fe-oxygen (O) system is known for its structural and magnetic complexity, rendering the generation of high-quality dataset challenging. In this work, we propose the Perturbation Augmented Space group structure Sampling (PASS) method to generate extensive and representative dataset consisting of small-cell structures with less than 10 atoms. We present a systematic approach to developing a first of its kind transferable machine learning interatomic potential (MLIP) for Fe-O system based on the atomic cluster expansion (ACE) framework. We thoroughly validate the accuracy and capability of the ACE MLIP across both pure Fe and Fe-O systems through bulk, surface, and interface properties. We showcase the formation of FeO-like structure in large-scale Fe oxidation simulation using the ACE MLIP. This work demonstrates that the PASS method yields an accurate and transferable MLIP which is capable of capturing the reactive complexity of oxide growth while remaining computationally practical for extended systems.

Figures

Figures reproduced from arXiv: 2607.28000 by the authors.

Figure 1
Figure 1. Overview of the PASS workflow for developing a transferable ACE MLIP for Fe-O system. (a) Dataset generation. Space group structures are first generated using PyXtal and then optimized through stepwise volume and full structural relaxation to obtain initial equilibrium configurations. Perturbed structures are subsequently generated from the relaxed structures through systematic rattle and stretch operations, introdu… view at source ↗
Figure 2
Figure 2. illustrates the diversity and representativeness of the final dataset for the Fe-O system generated following the PASS method [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Validation of the ACE MLIP for structural and mechanical properties of pure Fe. (a) Equation of state curves for body-centered cubic (BCC) and face-centered cubic (FCC) Fe predicted by the ACE MLIP and compared with DFT reference data. Ferromagnetic (FM) and antiferromagnetic (AFM) orderings are used in the corresponding DFT calculations for BCC and FCC Fe, respectively. (b) Temperature-dependent equilibrium lattice… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Validation of the ACE MLIP for thermodynamics and kinetic properties of Fe-O system. (a) Comparison of atomic energies between DFT calculations and ACE MLIP predictions for relaxation trajectories of Fe oxide bulks, surfaces, and Fe/FeO interface structures, assessing …
Figure 5
Figure 5. Figure 5: Large-scale BCC Fe surface oxidation simulation using the ACE MLIP. (a) Final configuration after 1 ns high-temperature oxidation of the BCC Fe (100) surface. Three regions are identified along the Z direction: an oxide layer near the surface (region Ⅰ), an intermediat…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

6 extracted references

  1. [8]

    Sun, Y . et al. Quan3ta3ve 3D evolu3on of colloidal nanopar3cle oxida3on in solu3on. Science (1979). 356, 303–307 (2017). 9. Tu, W. et al. In Situ Mul3scale Study of Iron Oxida3on at High Temperatures. Nano LeF. 25, 8089–8095 (2025). 10. Zhang, X., Zheng, P ., Ma, Y ., Jiang, Y . & Li, H. Atomic-scale understanding of oxida3on mechanisms of materials by c...

    1979

  2. [22]

    Thijs, L. C. et al. Effect of Fe–O ReaxFF on Liquid Iron Oxide Proper3es Derived from Reac3ve Molecular Dynamics. J. Phys. Chem. A 127, 10339–10355 (2023). 23. Wei, Z., Shuang, F. & Dey, P . From O adsorp3on to Fe oxide growth: Benchmarking reac3ve force fields and universal machine learning interatomic poten3als against DFT for BCC Fe surface oxida3on. Sur...

    2023

  3. [36]

    Liu, Y . et al. An automated framework for exploring and learning poten3al-energy surfaces. Nat. Commun. 16, 7666 (2025). 37. Poul, M., Huber, L., Bitzek, E. & Neugebauer, J. Systema3c atomic structure datasets for machine learning poten3als: Applica3on to defects in magnesium. Phys. Rev. B 107, 104103 (2023). 38. Poul, M., Huber, L. & Neugebauer, J. Auto...

    doi:10.48550/arxiv.2512.12843 2025

  4. [50]

    & Maresca, F

    Zhang, L., Csányi, G., van der Giessen, E. & Maresca, F. Efficiency, accuracy, and transferability of machine learning poten3als: Applica3on to disloca3ons and cracks in iron. Acta Mater. 270, 119788 (2024). 51. Peng, H. et al. Thermodynamics and kine3cs of martensi3c transforma3on in iron-based alloys via Bain path: Models and atomis3c simula3ons. Acta Mat...

    2024

  5. [62]

    Menon, S. et al. From electrons to phase diagrams with machine learning poten3als using pyiron based automated workflows. NPJ Comput. Mater. 10, 261 (2024). 63. Fredericks, S., Parrish, K., Sayre, D. & Zhu, Q. PyXtal: A Python library for crystal structure genera3on and symmetry analysis ✩,✩✩. Comput. Phys. Commun. 261, 107810 (2021). 64. De, S., Bartók, A...

    2024

  6. [76]

    & Drautz, R

    Lysogorskiy, Y ., Bochkarev, A., Mrovec, M. & Drautz, R. Ac3ve learning strategies for atomic cluster expansion models. Phys. Rev. Mater. 7, (2023). 77. Hjorth Larsen, A. et al. The atomic simula3on environment—a Python library for working with atoms. Journal of Physics: Condensed MaFer 29, 273002 (2017). 78. Bitzek, E., Koskinen, P ., Gähler, F., Moseler...

    2023

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.