Pith. sign in

REVIEW 4 major objections 8 minor 69 references

Coupled MC and MD with ML surrogates nails Fe-Co alloy transitions

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · glm-5.2

2026-07-09 10:16 UTC pith:XJ4X2ZRK

load-bearing objection Genuine methodological advance for coupled configurational-structural disorder in alloys, but the MC+MD coupling scheme has an unaddressed free-energy approximation that may bias the melting temperature prediction. the 4 major comments →

arxiv 2607.07456 v1 pith:XJ4X2ZRK submitted 2026-07-08 cond-mat.mtrl-sci

A Multi-Scale Machine Learning Framework for Coupled Chemical, Spin, and Structural Disorder in Alloys

Zhenyao Fang , Qimin Yan This is my paper
classification cond-mat.mtrl-sci
keywords alloysframeworkstructuralchemicaldisorderspinalloyfe-co
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper presents a computational framework that interleaves Monte Carlo sampling of chemical and spin configurations with molecular dynamics relaxation of atomic positions, using a graph neural network to evaluate energies and a machine learning interatomic potential to compute forces. The central claim is that this coupling is both necessary and sufficient to correctly capture the thermodynamics of alloys where chemical disorder, spin disorder, and structural distortions are simultaneously present. The authors demonstrate the framework on body-centered-cubic Fe-Co alloys with interstitial carbon. Without structural relaxation, the framework predicts an order-to-disorder transition at 780 K for pure Fe-Co and identifies a linear carbon-chain ground state. When structural distortions are included through NPT molecular dynamics, the predicted transition temperature shifts to 1,000 K, matching the experimental value of 1,006 K, and the predicted melting temperature of 1,690 K matches the experimental 1,700 K. The carbon-chain ground state disappears once lattice relaxation is allowed, because local distortions break the force-cancellation symmetry that stabilized it. The framework also predicts a tetragonal-to-nearly-cubic structural transition in Fe-Co-C as temperature increases and carbon disorders.

Core claim

The paper's central result is that configurational and structural degrees of freedom in disordered magnetic alloys are so strongly coupled that omitting structural relaxation from thermodynamic sampling produces qualitatively wrong predictions, not merely quantitatively imprecise ones. The shift from a 780 K to a 1,000 K transition temperature upon including structural distortions, the destruction of the carbon-chain ground state, and the emergence of a tetragonal-to-cubic structural transition all demonstrate that structural relaxation is a first-order effect on the thermodynamics. The framework achieves this by having the GNN evaluate energies on MD-relaxed snapshots, averaging several of,

What carries the argument

The framework operates as a loop: a Monte Carlo move (elemental swap, carbon hop, or spin flip) is proposed, a short NPT molecular dynamics trajectory relaxes the structure, several snapshots are extracted from that trajectory, the GNN evaluates the energy of each snapshot, and the average is used for Metropolis acceptance. The GNN takes atomic species and initial spin orientations as node features and predicts both total energy and self-consistently relaxed magnetic moments. The MLIP (fine-tuned CHGNet) computes forces and stresses through energy backpropagation but does not handle magnetic degrees of freedom, justified by the claim that magnetic configuration does not significantly affect,

Load-bearing premise

The framework uses a non-magnetic machine learning potential for structural dynamics while the GNN handles magnetic energy evaluation, justified by the claim that magnetic configuration does not significantly affect interatomic forces and stresses in this system. If that decoupling breaks down at higher temperatures or different compositions, the effective energies used for Metropolis acceptance would be biased.

What would settle it

If the predicted transition temperatures (1,000 K for ordering, 1,690 K for melting) are artifacts of the force field rather than genuine thermodynamic predictions, then applying the same framework to a different alloy system with known experimental transition temperatures should produce similarly accurate results. A single-system demonstration cannot distinguish a correct framework from a fortuitously parameterized one.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the framework generalizes beyond Fe-Co-C, it could predict phase diagrams and transition temperatures for high-entropy alloys where chemical, structural, and potentially magnetic disorder coexist, replacing expensive DFT-based sampling at a fraction of the cost.
  • The finding that structural distortions destroy the carbon-chain ground state suggests that earlier fixed-lattice predictions of ground-state configurations in interstitial-doped alloys may be systematically unreliable.
  • The tetragonal-to-cubic transition prediction for Fe-Co-C is directly testable by temperature-dependent X-ray or neutron diffraction experiments tracking lattice parameters.
  • The framework could be extended to compute magnetic anisotropy and magnetostriction as functions of composition and temperature by adding the relevant property as a GNN output.
  • The separation of roles between GNN (energy and magnetization) and MLIP (forces and stresses) offers a design pattern for other multi-property surrogate frameworks where no single model handles all required quantities.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 8 minor

Summary. This manuscript presents a computational framework that couples GNN-driven Monte Carlo sampling of configurational degrees of freedom (chemical swaps, spin flips) with MLIP-driven molecular dynamics for structural relaxation, applied to Fe-Co-C alloys. The GNN predicts total energies and site-resolved magnetic moments; a fine-tuned CHGNet provides forces and stresses. The authors report predicted order-to-disorder transition temperatures and melting temperatures in close agreement with experiment for Fe-Co, and predict tetragonal-to-nearly-cubic structural transitions in Fe-Co-C as temperature increases. The approach of separating energy evaluation (GNN, spin-aware) from force evaluation (CHGNet, non-magnetic) is a pragmatic design choice, and the comparison of results with and without structural distortions (Figs. 3 and 4) effectively demonstrates the necessity of including structural disorder.

Significance. The coupling of spin-aware GNN energy evaluation with MLIP-driven structural dynamics in a unified MC+MD framework is a genuinely useful methodological contribution for multi-disorder magnetic alloys, a class of systems that is difficult to treat with existing cluster expansion or classical force field approaches. The demonstration that structural distortions qualitatively change the predicted ground state (destabilizing the carbon chain) and quantitatively change the transition temperature is a clear and important result. The framework is general and transferable in principle. However, the precision of the headline numerical claims (1,000 K vs. 1,006 K; 1,690 K vs. 1,700 K) is not supported by any uncertainty quantification, which weakens the significance of the apparent agreement.

major comments (4)
  1. Methods, MC and MD Simulations section: The effective configurational energy used in the Metropolis acceptance criterion is the arithmetic mean of GNN energies over 5 MD snapshots, E_eff = (1/5) Σ_i E_GNN(R_i). When structural degrees of freedom are integrated out at finite temperature, the correct effective energy for the coarse-grained configurational MC is the configurational free energy F_config = -k_B T ln⟨exp(-β E_GNN)⟩_MD, not the arithmetic mean ⟨E_GNN⟩_MD. By Jensen's inequality, ⟨E⟩ ≥ F_config, with the difference equal to T S_structural. This bias is configuration-dependent and temperature-dependent; it is small when structural fluctuations are small (low T, rigid lattice) but could be substantial near melting where structural fluctuations are large. The manuscript does not discuss this approximation anywhere. This is load-bearing because the Metropolis acceptance criterion is
  2. Results and Discussions, Phase Transitions With Structural Distortions (Fig. 4): The predicted transition temperatures (1,000 K and 1,690 K) are reported without any error bars or uncertainty estimates. The simulations are stochastic (MC sampling with 2×10^4 steps per temperature, 5 MD snapshots per step), the GNN energy MAE is 34.0 meV/atom (comparable to k_B T ≈ 86 meV/atom at 1,000 K), and the 5 MD snapshots from a 1 ps trajectory with a Berendsen thermostat (τ_T = 10 fs) are highly correlated. Without uncertainty quantification—whether from block averaging, bootstrap resampling, or multiple independent runs—the claim of 'excellent agreement' with experimental values cannot be assessed. The closeness of the predictions to experiment (within 6 K and 10 K) may be partly fortuitous given these error sources.
  3. Results and Discussions, Phase Transitions With Structural Distortions: All simulations use a 3×3×3 BCC supercell (54 atoms for pure Fe-Co). No finite-size scaling analysis is presented. For order-to-disorder transitions, finite-size effects can shift transition temperatures by tens to hundreds of kelvin, and the sharpness of the heat capacity peak is directly affected by system size. The paper should at minimum discuss this limitation and its likely direction of bias, or show results for at least one additional cell size to demonstrate convergence.
  4. Results and Discussions, Training and Validation of GNN and MLIP Models: The GNN energy MAE of 34.0 meV/atom (1.821 eV/cell for 54 atoms) is large relative to the energy differences that govern MC acceptance. For context, the DFT energy difference between the two carbon interstitial sites C_i1 and C_i2 is reported as ~0.06 eV (§Phase Transitions Without Structural Distortions), which is far smaller than the per-cell MAE. The paper should discuss whether the GNN accuracy is sufficient to resolve the energy differences relevant to the MC sampling, and ideally provide an estimate of how the energy MAE propagates into uncertainty in the predicted transition temperatures.
minor comments (8)
  1. Fig. 1(a): The schematic is helpful but the connection between the MC accept/reject step and the MD trajectory could be clearer. Consider adding labels indicating the sequence of operations (propose move → MD → sample snapshots → average energies → Metropolis).
  2. §Methods, MC and MD Simulations: The heat capacity formula (Eq. 0.3) uses C_v, but the NPT-MD simulations sample an NPT ensemble. The relevant heat capacity at constant pressure would be C_p. Please clarify which ensemble quantity is being computed and whether the notation should be C_p.
  3. §Methods, MC and MD Simulations: The Berendsen thermostat and barostat are known to not generate a proper canonical or NPT ensemble (they do not reproduce the correct fluctuations). This is a known limitation; consider discussing whether a proper thermostat (e.g., Nosé-Hoover) would be more appropriate, or at minimum noting this as a limitation.
  4. §Results, Phase Transitions Without Structural Distortions: The Schottky anomaly estimate T ≈ 0.42 ΔE/k_B ≈ 292 K uses ΔE ≈ 0.06 eV. Please verify: 0.42 × 0.06 eV / k_B = 0.42 × 696 K ≈ 292 K. This is correct, but the factor 0.42 should be derived or cited more explicitly.
  5. Fig. 4(b): The anisotropic order parameters D_ab and D_c are defined in Eqs. (0.1) and (0.2), but the normalization by l̄ rather than a fixed reference lattice constant is only briefly justified. Consider elaborating on why this choice is appropriate for tracking symmetry changes independent of thermal expansion.
  6. §Results, Training and Validation: The GNN dataset is generated from only 500 MLIP snapshots (8 spin/chemical configurations each). It would be useful to report the chemical and compositional diversity of these 500 underlying structures, and whether the training set covers the configurations encountered during MC sampling at the relevant compositions (50:50 Fe-Co).
  7. The manuscript uses 'Wang-Landau sampling' in the framework description (§Results and Discussions, first paragraph) but all reported results appear to use Metropolis sampling. Please clarify whether Wang-Landau was used for any results, or remove the mention if it was only included as a general framework option.
  8. Several references appear to be from 2025-2026 (e.g., Refs. [2], [8], [47], [49], [50], [52]). Please verify these citations are to published or properly archived works.

Simulated Author's Rebuttal

4 responses · 3 unresolved

We thank the referee for a careful and constructive review. The referee raises four major points: (1) the use of arithmetic mean energy rather than configurational free energy in the Metropolis criterion, (2) absence of uncertainty quantification on transition temperatures, (3) lack of finite-size scaling analysis, and (4) whether the GNN energy MAE is sufficient to resolve relevant energy differences. We agree that points (1), (2), and (3) identify legitimate gaps in uncertainty treatment and discussion that we will address in revision. On point (4), we partially agree and will add discussion, but we also clarify why the per-atom MAE is not the relevant metric for the MC energy differences. We cannot fully resolve points (2) and (3) with new simulations in the revision timeframe but will add honest discussion of limitations and temper the headline claims accordingly.

read point-by-point responses

  1. Referee: The effective configurational energy used in the Metropolis acceptance criterion is the arithmetic mean of GNN energies over 5 MD snapshots, E_eff = (1/5) Σ E_GNN(R_i), rather than the configurational free energy F_config = -k_B T ln⟨exp(-β E_GNN)⟩_MD. By Jensen's inequality, ⟨E⟩ ≥ F_config, with the difference equal to T S_structural. This bias is configuration-dependent and temperature-dependent and is not discussed in the manuscript.

    Authors: The referee is correct that the arithmetic mean ⟨E_GNN⟩_MD is not the exact configurational free energy F_config = -k_B T ln⟨exp(-β E_GNN)⟩_MD, and that Jensen's inequality implies ⟨E⟩ ≥ F_config with the difference being T·S_structural. We acknowledge that this approximation is not discussed in the current manuscript, and we will add a dedicated discussion of this point in the revised Methods and Results sections. We note that the practical motivation for the arithmetic mean is computational: evaluating F_config requires exponentiating each of the 5 MD snapshots, which is feasible in principle but was not implemented. The bias is configuration-dependent and temperature-dependent, being smallest when structural fluctuations are small (low T, rigid lattice) and largest near melting. Importantly, the bias enters the Metropolis acceptance ratio as a difference between the proposed and current configurations, so if the structural entropy is similar across configurations at a given temperature, the bias partially cancels. Nevertheless, near melting where structural fluctuations differ dramatically between solid and liquid configurations, this cancellation may break down. We will add this discussion, characterize the approximation as a known limitation, and note that implementing the exact free energy formulation is a straightforward future improvement. We will also temper claims of quantitative accuracy near the melting transition accordingly. revision: partial

  2. Referee: The predicted transition temperatures (1,000 K and 1,690 K) are reported without error bars or uncertainty estimates. The simulations are stochastic, the GNN energy MAE is 34.0 meV/atom (comparable to k_B T at 1,000 K), and the 5 MD snapshots from a 1 ps trajectory with a Berendsen thermostat (τ_T = 10 fs) are highly correlated. Without uncertainty quantification, the claim of 'excellent agreement' cannot be assessed.

    Authors: We agree that the absence of uncertainty quantification is a significant gap, and we accept this criticism. We will address it in two ways in the revision. First, we will add bootstrap resampling of the MC trajectory to estimate statistical uncertainty on the heat capacity peak location, and we will report error bars on the predicted transition temperatures. Second, we will acknowledge the issue of snapshot correlation: 5 snapshots from a 1 ps trajectory with τ_T = 10 fs are indeed correlated, and we will discuss this as a source of potential bias in the effective energy estimate. We also agree that the GNN energy MAE of 34.0 meV/atom contributes systematic uncertainty that is not captured by bootstrap resampling of a single trajectory. We will therefore soften the language from 'excellent agreement' to 'good agreement within the limitations of the framework' and explicitly state that the closeness of the predictions to experimental values (within 6 K and 10 K) may be partly fortuitous given the combined error sources from GNN accuracy, snapshot correlation, finite system size, and the arithmetic mean approximation. We cannot run fully independent replica simulations within the revision timeframe, but we will add the bootstrap analysis and an honest error budget discussion. revision: partial

  3. Referee: All simulations use a 3×3×3 BCC supercell (54 atoms). No finite-size scaling analysis is presented. Finite-size effects can shift transition temperatures by tens to hundreds of kelvin, and the sharpness of the heat capacity peak is directly affected by system size.

    Authors: We agree that finite-size effects are a legitimate concern and that the absence of any discussion of this limitation is a gap in the manuscript. We will add a dedicated paragraph discussing finite-size effects. We note that for the B2 order-disorder transition in Fe-Co, the transition is expected to be first-order or weakly first-order, for which finite-size shifts are typically smaller than for continuous transitions but still non-negligible. The 54-atom supercell is admittedly small, and we expect the transition temperature to be shifted relative to the thermodynamic limit and the heat capacity peak to be broadened. We cannot run additional cell sizes within the revision timeframe due to the computational cost of the coupled MC+MD protocol (each MC step requires 5 GNN energy evaluations plus a 1 ps MD trajectory), but we will explicitly state this limitation, discuss the likely direction of bias, and note that finite-size scaling is a necessary future investigation. We will also adjust the manuscript language to acknowledge that the quantitative agreement with experiment may benefit from error cancellation between finite-size effects and other error sources. revision: partial

  4. Referee: The GNN energy MAE of 34.0 meV/atom (1.821 eV/cell for 54 atoms) is large relative to the energy differences that govern MC acceptance, such as the ~0.06 eV DFT energy difference between C_i1 and C_i2 interstitial sites. The paper should discuss whether the GNN accuracy is sufficient to resolve the relevant energy differences.

    Authors: We partially agree with this concern and will add discussion, but we also note an important distinction. The per-cell MAE of 1.821 eV is computed over the full dataset spanning the entire composition range (0-100% Fe, 0-11.1% C) and all magnetic configurations, including configurations far from equilibrium. The relevant quantity for MC acceptance is not the absolute MAE but the error on energy differences between configurations that differ by a single swap or spin flip, which are typically much closer in composition and energy. We will add this clarification and, where possible, report the MAE on energy differences between closely related configurations (e.g., single-swap pairs) as a more relevant accuracy metric. That said, we acknowledge that for the C_i1 vs C_i2 energy difference of ~0.06 eV, the GNN accuracy on such small differences is a legitimate concern, and we will discuss this honestly. We note that the MC results for the two-level carbon system (Fig. 3, middle panel) show a peak at temperatures significantly above the Schottky anomaly prediction, which is consistent with the Fe-Co coupling effect described in the manuscript; if the GNN were unable to resolve the C_i1-C_i2 energy difference at all, this feature would not be reproduced. We will add a discussion of error propagation from GNN energy errors to transition temperature uncertainty, acknowledging that a full quantitative error propagation analysis is beyond the scope of the current work. revision: partial

standing simulated objections not resolved
  • We cannot provide full finite-size scaling results (multiple supercell sizes) within the revision timeframe due to the computational cost of the coupled MC+MD protocol.
  • We cannot provide results from multiple fully independent MC+MD replica runs within the revision timeframe; bootstrap resampling of existing trajectories is the best we can offer for statistical uncertainty estimation.
  • Implementing the exact configurational free energy F_config = -k_B T ln⟨exp(-β E_GNN)⟩_MD in place of the arithmetic mean requires code modifications and re-running all simulations, which is not feasible in the revision timeframe; we will discuss the approximation and its limitations but cannot replace it with the exact formulation.

Circularity Check

0 steps flagged

No significant circularity; self-citations are contextual, not load-bearing

full rationale

The paper's central predictions (order-to-disorder transition temperature of 1,000 K and melting temperature of 1,690 K) are emergent outputs of MC+MD sampling driven by models trained on DFT energies, forces, stresses, and magnetic moments. No parameter is fitted to the experimental transition temperatures and then presented as a prediction. The GNN is trained on DFT total energies and relaxed magnetic moments (4,000 configurations), and the CHGNet MLIP is fine-tuned on DFT forces and stresses (2,500 snapshots) — neither target is the thermodynamic property being predicted. The self-citations (Ref [10], Fang & Yan 2024; Ref [47], Fang et al. 2026; Ref [54], Fang & Yan 2025) provide methodological context for GNN-driven MC and atomic representations, but the current paper fully re-describes its methodology and does not invoke any self-cited uniqueness theorem or ansatz as load-bearing for its central claims. The skeptic's concern about using the arithmetic mean of GNN energies rather than the configurational free energy F = -kT ln⟨exp(-βE)⟩ is a correctness/approximation risk, not a circularity issue — the paper is not defining its effective energy in terms of the quantity it claims to predict. The derivation chain is self-contained against external experimental benchmarks, and the self-citations do not reduce the central results to their inputs by construction. Score 1 reflects the presence of minor self-citations that are not load-bearing for the paper's claims.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 0 invented entities

The framework introduces no new physical entities, particles, or forces. The GNN and MLIP are standard machine learning architectures applied to DFT training data. The free parameters are standard DFT+U values and ML hyperparameters. The key axioms are domain assumptions about the decoupling of magnetic state from structural forces and the adequacy of the supercell size, both of which are load-bearing for the central claims.

free parameters (7)
  • U_Fe (DFT+U Hubbard parameter) = 4 eV
    Standard but empirically chosen value for Fe 3d orbitals; not derived from first principles in this paper.
  • U_Co (DFT+U Hubbard parameter) = 3 eV
    Standard but empirically chosen value for Co 3d orbitals; not derived from first principles in this paper.
  • GNN hyperparameters (layers, channels, learning rate, weight decay, batch size) = Optimized via Bayesian optimization (Optuna)
    Hyperparameters selected to minimize validation loss; not physically motivated.
  • MLIP fine-tuning hyperparameters (learning rate, batch size, trainable modules) = Optimized via Bayesian optimization (Optuna)
    Hyperparameters selected to minimize validation loss.
  • MD timestep = 2 fs
    Standard choice; not justified for this system specifically.
  • Number of MD snapshots per MC step = 5
    Chosen for energy averaging; no convergence study provided.
  • MC steps per temperature = 1e5 (no distortion) or 2e4 (with distortion)
    Chosen values; no convergence analysis on thermodynamic properties presented.
axioms (4)
  • domain assumption Magnetic configurations do not significantly affect interatomic forces and stresses in Fe-Co-C alloys
    Invoked in the main text to justify using non-magnetic CHGNet for structural dynamics; supported only by reference to Supplementary Information 1.3.
  • domain assumption 3x3x3 supercell (54 atoms) is sufficient to capture the thermodynamic phase transitions
    Used throughout all MC simulations; finite-size effects are not discussed.
  • standard math DFT+U with PBE functional provides accurate reference energies for Fe-Co-C alloys
    Standard computational chemistry assumption; the choice of U values affects training data quality.
  • ad hoc to paper Averaging GNN energies over 5 MD snapshots per MC step yields an unbiased effective configurational energy
    The framework defines the effective energy this way (Methods section); no proof or convergence study that this averaging is sufficient.

pith-pipeline@v1.1.0-glm · 17128 in / 3181 out tokens · 482051 ms · 2026-07-09T10:16:11.405541+00:00 · methodology

0 comments
read the original abstract

Understanding the thermodynamic properties of disordered magnetic alloys requires a unified treatment of configurational (chemical, spin, etc.) and structural degrees of freedom, which has remained beyond the scope of existing computational frameworks. Here we present a general framework that integrates machine learning models (such as graph neural networks and machine learning interatomic potentials) and statistical sampling methods (such as Monte Carlo and molecular dynamics simulations) to study the coupled chemical, spin, and structural disorder in alloys. We demonstrate the framework on body-centered-cubic Fe-Co alloys with interstitial carbon dopants, where the Fe-Co host exhibits intrinsic chemical and spin disorder, and the interstitial carbon introduces additional structural disorder through local lattice distortions, making the system a prototypical multi-disorder magnetic alloy. The framework predicts the order-to-disorder phase transition temperature and the melting temperature of Fe-Co alloy to be 1,000 K and 1,690 K, in excellent agreement with the experimentally measured values of 1,006 K and approximately 1,700 K, respectively. It also predicts the tetragonal-to-nearly-cubic structural transitions in Fe-Co-C alloy as temperature increases. These results establish the framework as a reliable tool for studying multi-disorder alloys, with applications to complex disordered systems such as high-entropy alloys, multiferroics, and spintronic devices.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

69 extracted references · 69 canonical work pages · 5 internal anchors

  1. [1]

    Bean, C. P. & Rodbell, D. S. Magnetic disorder as a first-order phase transfor- mation.Phys. Rev.126, 104–115 (1962). URL https://link.aps.org/doi/10.1103/ PhysRev.126.104

    work page 1962

  2. [2]

    L., Silevitch, D

    Armstrong, S. L., Silevitch, D. M. & Rosenbaum, T. F. Experimental conse- quences of disorder at an antiferromagnetic quantum phase transition.Nature Communications16, 2178 (2025)

    work page 2025

  3. [3]

    Pei, Z.et al.Theory-guided design of high-entropy alloys with enhanced strength- ductility synergy.Nature Communications14, 2519 (2023)

    work page 2023

  4. [4]

    & Wei, S

    Zhu, C., Xu, L., Liu, M., Guo, M. & Wei, S. A review on improving mechanical properties of high entropy alloy: interstitial atom doping.Journal of Materials Research and Technology24, 7832–7851 (2023)

    work page 2023

  5. [5]

    Dasari, S.et al.Exceptional enhancement of mechanical properties in high- entropy alloys via thermodynamically guided local chemical ordering.Proceedings of the National Academy of Sciences120, e2211787120 (2023)

    work page 2023

  6. [6]

    Anomalous Dirac point transport due to extended defects in bilayer graphene

    Shallcross, S., Sharma, S. & Weber, H. B. Anomalous Dirac point transport due to extended defects in bilayer graphene.Nature Communications8, 342 (2017). 1612.01472

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  7. [7]

    Anomalous transport due to Weyl fermions in the chiral antiferromagnets Mn$_3$$X$, $X$ = Sn, Ge

    Chen, T.et al.Anomalous transport due to Weyl fermions in the chiral anti- ferromagnets Mn3X, X = Sn, Ge.Nature Communications12, 572 (2021). 2011.10942

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  8. [8]

    Fang, H.et al.Different charge transport mechanisms in ti3c2tx mxene monoflakes and multiflakes.The Journal of Physical Chemistry Letters16, 7515–7521 (2025)

    work page 2025

  9. [9]

    J., Pan, J., Tamboli, A

    Cordell, J. J., Pan, J., Tamboli, A. C., Tucker, G. J. & Lany, S. Probing config- urational disorder inzngen2 using cluster-based monte carlo.Phys. Rev. Mater. 5, 024604 (2021)

    work page 2021

  10. [10]

    & Yan, Q

    Fang, Z. & Yan, Q. Towards accurate prediction of configurational disorder prop- erties in materials using graph neural networks.npj Computational Materials10, 91 (2024)

    work page 2024

  11. [11]

    J., Dove, M

    Cliffe, M. J., Dove, M. T., Drabold, D. A. & Goodwin, A. L. Structure determi- nation of disordered materials from diffraction data.Phys. Rev. Lett.104, 125501 (2010)

    work page 2010

  12. [12]

    & Wei, S.-H

    Yang, J., Wang, J., Yang, C., Zhang, W. & Wei, S.-H. First-principles study of the order-disorder transition and its effects on the optoelectronic property of aBis2 (a= Na, k).Phys. Rev. Mater.4, 085402 (2020)

    work page 2020

  13. [13]

    & Salmon, P.Molecular dynamics sim- ulations of disordered materials: from network glasses to phase-change memory alloys

    Massobrio, C., Du, J., Bernasconi, M. & Salmon, P.Molecular dynamics sim- ulations of disordered materials: from network glasses to phase-change memory alloys. Springer series in materials science ; volume 215 (Springer, Cham, 2015)

    work page 2015

  14. [14]

    Springer series in solid-state sciences, 175 (Springer, Berlin, 2012)

    Kakehashi, Y.Modern theory of magnetism in metals and alloys. Springer series in solid-state sciences, 175 (Springer, Berlin, 2012). URL https://link.ezproxy. neu.edu/login?url=https://link.springer.com/10.1007/978-3-642-33401-6

    work page doi:10.1007/978-3-642-33401-6 2012

  15. [15]

    & Nakamura, Y

    Kaneyoshi, T. & Nakamura, Y. Magnetic properties of disordered Ising ternary alloys.Journal of Physics: Condensed Matter10, 5359 (1998)

    work page 1998

  16. [16]

    S.-D., Bolarín-Miró, A

    Jesús, F. S.-D., Bolarín-Miró, A. M., Escobedo, C. A. C., Torres-Villaseñor, G. 15 & Vera-Serna, P. Structural Analysis and Magnetic Properties of FeCo Alloys Obtained by Mechanical Alloying.Journal of Metallurgy2016, 1–8 (2016)

    work page 2016

  17. [17]

    Sundar, R. S. & Deevi, S. C. Soft magnetic FeCo alloys: alloy development, processing, and properties.International Materials Reviews50, 157–192 (2005)

    work page 2005

  18. [18]

    & Heinonen, O

    Burkert, T., Nordström, L., Eriksson, O. & Heinonen, O. Giant Magnetic AnisotropyinTetragonalFeCoAlloys.Physical Review Letters93,027203(2004)

    work page 2004

  19. [19]

    Near equiatomic FeCo alloys: Constitution, mechanical and magnetic properties.Progress in Materials Science50, 816–880 (2005)

    Sourmail, T. Near equiatomic FeCo alloys: Constitution, mechanical and magnetic properties.Progress in Materials Science50, 816–880 (2005)

    work page 2005

  20. [20]

    Koutsopoulos, S., Barfod, R., Eriksen, K. M. & Fehrmann, R. Synthesis and characterization of iron-cobalt (FeCo) alloy nanoparticles supported on carbon. Journal of Alloys and Compounds725, 1210–1216 (2017)

    work page 2017

  21. [21]

    Torchio, R.et al.Pressure-induced structural and magnetic phase transitions in ordered and disordered equiatomic FeCo.Physical Review B88, 184412 (2013)

    work page 2013

  22. [22]

    Schoen, M. A. W.et al.Ultra-low magnetic damping of a metallic ferromagnet. Nature physics12, 839–842 (2016)

    work page 2016

  23. [23]

    Kinetics of short-range and long-range B2 ordering in FeCo.Physical Review B44, 9805–9811 (1991)

    Fultz, B. Kinetics of short-range and long-range B2 ordering in FeCo.Physical Review B44, 9805–9811 (1991)

    work page 1991

  24. [24]

    Seehra, M. S. & Silinsky, P. Order-disorder andα-γtransitions in FeCo.Physical Review B13, 5183–5187 (1976)

    work page 1976

  25. [25]

    Oyedele, J. A. & Collins, M. F. Composition dependence of the order-disorder transition in iron-cobalt alloys.Physical Review B16, 3208–3212 (1977)

    work page 1977

  26. [26]

    Wells, S.et al.A study of Fe-B and Fe-Co-B alloy particles produced by reduction with borohydride.Journal of Physics: Condensed Matter1, 8199 (1989)

    work page 1989

  27. [27]

    W., Wells, S., Meagher, A., Mo/rup, S

    Charles, S. W., Wells, S., Meagher, A., Mo/rup, S. & Wonterghem, J. v. Prop- erties of amorphous FeCoB alloy particles (abstract).Journal of Applied Physics 64, 5508–5508 (1988)

    work page 1988

  28. [28]

    & Werwiński, M

    Marciniak, W. & Werwiński, M. Structural and magnetic properties of Fe-Co-C alloys with tetragonal deformation: A first-principles study.Physical Review B 108, 214433 (2023). 2307.05709

    work page arXiv 2023

  29. [29]

    & Gedanken, A

    Holodelshikov, E., Perelshtein, I. & Gedanken, A. Synthesis of Air Stable FeCo/C Alloy Nanoparticles by Decomposing a Mixture of the Corresponding Metal- Acetyl Acetonates under Their Autogenic Pressure.Inorganic Chemistry50, 1288–1294 (2011)

    work page 2011

  30. [30]

    K.et al.Stabilization of the tetragonal distortion of fexco1−x alloys by c impurities: A potential new permanent magnet.Phys

    Delczeg-Czirjak, E. K.et al.Stabilization of the tetragonal distortion of fexco1−x alloys by c impurities: A potential new permanent magnet.Phys. Rev. B89, 144403 (2014). URL https://link.aps.org/doi/10.1103/PhysRevB.89.144403

    work page doi:10.1103/physrevb.89.144403 2014

  31. [31]

    & Hong, J

    Khan, I. & Hong, J. Magnetic anisotropy of c and n doped bulk feco alloy: A first principles study.Journal of magnetism and magnetic materials388, 101–105 (2015)

    work page 2015

  32. [32]

    Effect of disorder on transport properties in a tight-binding model for lead halide perovskites

    Ashhab,S.,Voznyy,O.,Hoogland,S.,Sargent,E.H.&Madjet,M.E. Effectofdis- order on transport properties in a tight-binding model for lead halide perovskites. Scientific Reports7, 8902 (2017). 1703.03574

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  33. [33]

    Princeton series in physics (Princeton University Press, Princeton, New Jersey, 1986 - 1986)

    Chowdhury, D.Spin glasses and other frustrated systems. Princeton series in physics (Princeton University Press, Princeton, New Jersey, 1986 - 1986). URL http://ebookcentral.proquest.com/lib/northeastern-ebooks/detail. action?docID=3030635. 16

    work page 1986

  34. [34]

    Atomic cluster expansion for accurate and transferable interatomic potentials.Phys

    Drautz, R. Atomic cluster expansion for accurate and transferable interatomic potentials.Phys. Rev. B99, 014104 (2019)

    work page 2019

  35. [35]

    Sanchez, J. M. Cluster expansions and the configurational energy of alloys.Phys. Rev. B48, 14013–14015 (1993)

    work page 1993

  36. [36]

    & Ong, S

    Chen, C., Ye, W., Zuo, Y., Zheng, C. & Ong, S. P. Graph networks as a universal machine learning framework for molecules and crystals.Chem. Mater.31, 3564– 3572 (2019)

    work page 2019

  37. [37]

    & Sumpter, B

    Fung, V., Zhang, J., Juarez, E. & Sumpter, B. G. Benchmarking graph neural networks for materials chemistry.npj Comput. Mater.7, 84 (2021)

    work page 2021

  38. [38]

    Reiser, P.et al.Graph neural networks for materials science and chemistry. Commun. Mater.3, 93 (2022)

    work page 2022

  39. [39]

    Wang, G.et al.Machine learning interatomic potential: Bridge the gap between small-scale models and realistic device-scale simulations.iScience27, 109673 (2024)

    work page 2024

  40. [40]

    A practical guide to machine learning interatomic potentials -- Status and future

    Jacobs, R.et al.A practical guide to machine learning interatomic potentials – Status and future.Current Opinion in Solid State and Materials Science35, 101214 (2025). 2503.09814

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  41. [41]

    Muralles, M., Oh, J. T. & Chen, Z. Molecular dynamics study of FeCo phase transitions and thermal properties based on an improved 2NN MEAM potential. Journal of Materials Research and Technology19, 1102–1110 (2022)

    work page 2022

  42. [42]

    & Löser, W

    Woodcock, T., Hermann, R. & Löser, W. Development of a metastable phase diagram to describe solidification in undercooled Fe–Co melts.Calphad31, 256– 263 (2007)

    work page 2007

  43. [43]

    Rodriguez, J. E. & Matson, D. M. Thermodynamic modeling of the solidification path of levitated Fe–Co alloys.Calphad49, 87–100 (2015)

    work page 2015

  44. [44]

    & Bellaiche, L

    Prokhorenko, S., Kalke, K., Nahas, Y. & Bellaiche, L. Large scale hybrid monte carlo simulations for structure and property prediction.npj Comput. Mater.4, 80 (2018)

    work page 2018

  45. [45]

    & Landau, D

    Wang, F. & Landau, D. P. Efficient, multiple-range random walk algorithm to calculate the density of states.Phys. Rev. Lett.86, 2050–2053 (2001)

    work page 2050

  46. [46]

    C., Torbrügge, S

    Zhou, C., Schulthess, T. C., Torbrügge, S. & Landau, D. P. Wang-landau algo- rithm for continuous models and joint density of states.Phys. Rev. Lett.96, 120201 (2006)

    work page 2006

  47. [47]

    & Yan, Q

    Fang, Z., Hsu, T.-W. & Yan, Q. Learning atomic representations for data-driven materials design.AI for Science2, 013001 (2026)

    work page 2026

  48. [48]

    Deng, B.et al.Chgnet as a pretrained universal neural network potential for charge-informed atomistic modelling.Nature Machine Intelligence1–11 (2023)

    work page 2023

  49. [49]

    Trestman, M., Gugler, S., Faber, F. A. & Lilienfeld, O. A. v. Gradient guided fur- thest point sampling for robust training set selection.Machine Learning: Science and Technology7, 035047 (2026). 2510.08906

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  50. [50]

    Yang, Z.et al.Efficient equivariant model for machine learning interatomic potentials.npj computational materials11, 49–10 (2025)

    work page 2025

  51. [51]

    InProceedings of the Thirtieth International Joint Conference on Artificial Intelligence(InternationalJointConferencesonArtificial 17 Intelligence, 2021)

    Shi, Y.et al.Masked label prediction: Unified message passing model for semi- supervised classification. InProceedings of the Thirtieth International Joint Conference on Artificial Intelligence(InternationalJointConferencesonArtificial 17 Intelligence, 2021)

    work page 2021

  52. [52]

    Xu, W.et al.Spin-informed universal graph neural networks for simulating mag- netic ordering.Proceedings of the National Academy of Sciences - PNAS122, e2422973122– (2025)

    work page 2025

  53. [53]

    Jiang, Y.et al.Topological representations of crystalline compounds for the machine-learning prediction of materials properties.npj Computational Materials 7, 28 (2021)

    work page 2021

  54. [54]

    & Yan, Q

    Fang, Z. & Yan, Q. Leveraging persistent homology features for accurate defect formation energy predictions via graph neural networks.Chemistry of materials 37, 1531–1540 (2025)

    work page 2025

  55. [55]

    Batzner, S.et al.E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials.Nature communications13, 2453–2453 (2022)

    work page 2022

  56. [56]

    Tari, A.The specific heat of matter at low temperatures(Imperial College Press, London, 2003)

    work page 2003

  57. [57]

    L., Tchougréeff, A

    Deringer, V. L., Tchougréeff, A. L. & Dronskowski, R. Crystal orbital hamilton population (cohp) analysis as projected from plane-wave basis sets.The Journal of Physical Chemistry A115, 5461–5466 (2011)

    work page 2011

  58. [58]

    T., Salje, E

    Eckstein, J. T., Salje, E. K. H., Howard, C. J. & Carpenter, M. A. Symmetry and strain analysis of combined electronic and structural instabilities in tungsten trioxide, W O 3.Journal of Applied Physics131, 215101 (2022)

    work page 2022

  59. [59]

    A., McKnight, R

    Carpenter, M. A., McKnight, R. E. A., Howard, C. J. & Knight, K. S. Symmetry andstrainanalysisofstructuralphasetransitionsinPr0.48Ca0.52MnO3.Physical Review B82, 094101 (2010)

    work page 2010

  60. [60]

    A., Salje, E

    Carpenter, M. A., Salje, E. K. H. & Graeme-Barber, A. Spontaneous strain as a determinant of thermodynamic properties for phase transitions in minerals. European Journal of Mineralogy10, 621–691 (1998)

    work page 1998

  61. [61]

    & Hafner, J

    Kresse, G. & Hafner, J. Ab initio molecular dynamics for liquid metals.Phys. Rev. B47, 558–561 (1993)

    work page 1993

  62. [62]

    & Furthmüller, J

    Kresse, G. & Furthmüller, J. Efficiency of ab-initio total energy calculations for metals and semiconductors using a plane-wave basis set.Comput. Mater. Sci.6, 15–50 (1996)

    work page 1996

  63. [63]

    & Joubert, D

    Kresse, G. & Joubert, D. From ultrasoft pseudopotentials to the projector augmented-wave method.Phys. Rev. B59, 1758–1775 (1999)

    work page 1999

  64. [64]

    Blöchl, P. E. Projector augmented-wave method.Phys. Rev. B50, 17953–17979 (1994)

    work page 1994

  65. [65]

    P., Burke, K

    Perdew, J. P., Burke, K. & Ernzerhof, M. Generalized gradient approximation made simple.Phys. Rev. Lett.77, 3865–3868 (1996)

    work page 1996

  66. [66]

    I., Zaanen, J

    Anisimov, V. I., Zaanen, J. & Andersen, O. K. Band theory and mott insulators: Hubbard u instead of stoner i.Phys. Rev. B44, 943–954 (1991). URL https: //link.aps.org/doi/10.1103/PhysRevB.44.943

    work page doi:10.1103/physrevb.44.943 1991

  67. [67]

    & de Gironcoli, S

    Cococcioni, M. & de Gironcoli, S. Linear response approach to the calculation of the effective interaction parameters in theLDA + Umethod.Phys. Rev. B71, 035105 (2005). URL https://link.aps.org/doi/10.1103/PhysRevB.71.035105

    work page doi:10.1103/physrevb.71.035105 2005

  68. [68]

    & Koyama, M

    Akiba, T., Sano, S., Yanase, T., Ohta, T. & Koyama, M. Optuna: A next- generation hyperparameter optimization framework. InProceedings of the 18 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, ACM, 2019, pp 2623–2631

    work page 2019

  69. [69]

    Berendsen, H. J. C., Postma, J. P. M., Gunsteren, W. F. v., DiNola, A. & Haak, J. R. Molecular dynamics with coupling to an external bath.The Journal of Chemical Physics81, 3684–3690 (1984). 19

    work page 1984