Pith. sign in

REVIEW 4 major objections 6 minor 55 references

Universal machine-learning potentials, used without retraining, can map the energy of carbon, nitrogen, oxygen, and hydrogen at thousands of interstitial sites in a titanium-niobium gum-metal alloy, revealing that titanium-rich surroundings

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 · deepseek-v4-flash

2026-08-03 18:19 UTC pith:KVBHCWLG

load-bearing objection Useful large-scale uMLIP comparison for interstitial energetics in a gum metal alloy, but the missing H DFT validation weakens the central transferability claim. the 4 major comments →

arxiv 2512.05568 v3 pith:KVBHCWLG submitted 2025-12-05 cond-mat.mtrl-sci

Revealing interstitial energetics in Ti-23Nb-0.7Ta-2Zr gum metal base alloy via universal machine learning interatomic potentials

Miroslav Lebeda , Jan Drahokoupil , Veronika Maz\'a\v{c}ov\'a , Petr Vl\v{c}\'ak This is my paper
classification cond-mat.mtrl-sci
keywords universal machine-learning interatomic potentialsinterstitial energeticsgum metal alloyTi-Nb-Ta-Zrspecial quasirandom structureshydrogen site preferencedefect chemistryDFT validation
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.

The paper sets out to show that three off-the-shelf, universally pretrained machine-learning interatomic potentials can stand in for DFT when mapping the energetics of light interstitial atoms (C, N, O, H) in a disordered Ti-23Nb-0.7Ta-2Zr gum-metal alloy. Using those potentials to relax and score 6,750 interstitial configurations in three 250-atom special quasirandom structures, it identifies two dominant chemical rules: the more titanium neighbors an interstitial has, the lower its energy, and the closer niobium sits, the higher its energy; tantalum and zirconium show no statistically significant effect. It also finds that the body-centered cubic lattice's geometric preferences survive chemical disorder—C, N, and O settle into octahedral sites while H prefers tetrahedral sites—for two of the three potentials, with the third (SevenNet-0) placing H in octahedral coordination, which the authors treat as a model limitation. If these findings hold, defect-energetics surveys that were previously too expensive for first-principles methods become routine, giving alloy designers a practical screening rule for interstitial behavior. The direct DFT check reported in the main text is limited to six oxygen configurations; broader claims about C, N, and H rest on the transferability of the universal potentials.

Core claim

On its own terms, the discovery is that a suite of universal machine-learning potentials—MACE-MATPES-PBE-0, Orb-v3, and SevenNet-0—can, without system-specific training, reproduce the energetic hierarchy among interstitial configurations in a chemically disordered bcc alloy. Across 6,750 relaxed configurations, every potential predicts a wide energy spread of roughly 1–3 eV that tracks local chemistry: increasing Ti nearest-neighbor count lowers energy, decreasing Nb–interstitial distance raises it, and Zr and Ta are statistically invisible. Two of the potentials (MACE and Orb-v3) preserve the known bcc site preferences, relaxing C, N, and O into octahedral positions and H into tetrahedral p

What carries the argument

The central objects are three universal machine-learning interatomic potentials (uMLIPs): pretrained neural-network models that output energies and forces for arbitrary atomic configurations without being fitted to this alloy. They replace DFT in the workflow, relaxing every candidate interstitial site and assigning it an energy. The statistical engine is a two-descriptor correlation analysis—each interstitial is characterized by the nearest-neighbor count of each host element and the minimum distance to each species—and Pearson correlation coefficients translate those descriptors into the paper's chemical-trend claims. A small DFT benchmark on six oxygen configurations serves as the calibra

Load-bearing premise

The load-bearing premise is that the three pretrained universal potentials, never trained on this alloy or its interstitial solutes, produce a faithful energy ranking for C, N, O, and H in this disordered alloy; the only direct first-principles check in the main text covers six oxygen configurations, so if any potential mis-ranks C, N, or H sites—or if the unreported H-DFT result contradicts the abstract—the chemical trends and the SevenNet-0 limitation claim would be unsuppo

What would settle it

Take the same three 250-atom SQS supercells and relax a stratified sample of C, N, and H interstitial configurations spanning the energy range each uMLIP predicts (about ten per element), using DFT with the paper's settings; compare the DFT energetic ordering and final site (octahedral vs tetrahedral) to the uMLIP results. If DFT reverses the relative energies of multiple configurations, or puts H in tetrahedral sites for SevenNet-0's octahedral minima, the paper's transferability thesis and its hydrogen-specific limitation claim would not survive.

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

If this is right

  • Interstitial design in gum-metal-type alloys can be guided by local chemistry: enriching the matrix in Ti should lower the energy of dissolved C, N, O, and H, while Nb-rich regions should repel them.
  • The persistence of bcc site preferences across chemical disorder means C, N, and O can be assumed to sit in octahedral sites and H in tetrahedral sites in this alloy class, simplifying thermodynamic and diffusion models.
  • uMLIP-relaxed geometries are good starting structures for DFT, reducing subsequent quantum-mechanical relaxation times from many hours to routine runs, making validation of large configurational sets affordable.
  • The same workflow is directly portable to other multicomponent alloys and other point defects, making thousand-configuration statistical surveys of defect energetics feasible without HPC access.
  • Cross-model agreement can serve as a built-in quality filter: when two independently trained universal potentials agree on a trend, confidence is higher; when they disagree (as with SevenNet-0 and H), targeted DFT validation is indicated.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper only benchmarks DFT against O configurations; a natural next calculation is a DFT set of C, N, and H configurations spanning each model's full energy range, which would put the universal-transferability claim on much firmer ground.
  • The Ti-attracts / Nb-repels rule, if confirmed by broader DFT, suggests an alloy-design lever: local composition fluctuations or phase separation into Ti-rich and Nb-rich regions should redistribute interstitials spatially, potentially controlling oxygen strengthening or hydrogen embrittlement in gum-metal components.
  • The SevenNet-0 hydrogen discrepancy points toward an element-specific blind-spot pattern in universal potentials: a cheap pre-screening check would be to compare a foundation model's predicted site preference against known bcc crystal chemistry, which would have caught the H error without any DFT.
  • The statistical protocol itself—three independent SQS cells, 2,250 interstitial sites per element per potential—could serve as a reusable benchmark for evaluating future universal potentials, exposing both energy distributions and site-preference errors more sharply than single-configuration tests.

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 / 6 minor

Summary. The paper applies three universal machine-learning interatomic potentials (MACE-MATPES-PBE-0, Orb-v3, SevenNet-0) to relax and score 6,750 octahedral and tetrahedral interstitial configurations (C, N, O, H) in three 250-atom SQS representations of the Ti-23Nb-0.7Ta-2Zr gum-metal base alloy. It reports broad energy distributions, site-preference differences between models (C/N/O in octahedral sites; H in tetrahedral sites for MACE and Orb, but octahedral for SevenNet-0), and Pearson correlations showing Ti-rich coordination is stabilizing while close Nb is destabilizing, with negligible Zr/Ta influence. A DFT benchmark of six O configurations is used to support the claim that the uMLIPs capture the energetic ordering without system-specific training. The paper also introduces an open-source GUI, uMLIP-Interactive, for running uMLIP calculations.

Significance. If the uMLIP energies are faithful proxies for DFT across the four interstitial species, the study would provide a statistically broad and computationally efficient map of interstitial energetics in a relevant gum-metal alloy, with potential guidance for alloy design. The paper's strengths include systematic sampling of thousands of configurations, use of three independent universal potentials, public release of scripts and the GUI, and an explicit (though narrow) DFT benchmark. However, the empirical support is considerably narrower than the central claim: the only direct DFT validation covers six O configurations, no DFT checks for C, N, or H appear, and the abstract's assertion that DFT confirms H's tetrahedral preference is not present in the main text. The 'statistically significant' language for the Zr/Ta null result is also unsupported by significance testing. These gaps are load-bearing because the paper's main conclusion is uMLIP transferability across chemically distinct interstitial environments.

major comments (4)
  1. [Abstract; §3.3; §4.1] The abstract claims that 'DFT validation confirms that tetrahedral configurations are energetically more favorable than octahedral sites for H interstitials,' but no H DFT calculation is reported in the main text or supplementary information. The only DFT validation in §3.3 is for six O configurations. Since the H site preference is central to the paper's assessment of SevenNet-0 (§3.1, §4.1), this claim must be substantiated with DFT data or removed/qualified. As written, the abstract overstates the evidence.
  2. [§4.3; §3.3] The conclusion that 'uMLIPs can capture the energetic hierarchy of interstitials in the Ti-23Nb-0.7Ta-2Zr alloy reasonably well' is generalized to C, N, O, and H, but the only direct DFT benchmark is six O configurations. No DFT validation for C, N, or H is presented. Because transferability across interstitial species is precisely the load-bearing claim, the authors should either add DFT benchmarks for at least H (and ideally C and N) or restrict the fidelity conclusion to O, presenting the other elements' trends as model predictions requiring further validation.
  3. [§3.2] The statement that Zr and Ta show 'no statistically significant influence' is not supported by any significance test. The manuscript uses a heuristic threshold |r| ≤ 0.1 for 'negligible' correlations and does not report p-values or confidence intervals. With n = 6750, even very small correlations can be statistically significant. Please report significance measures or rephrase the claim to describe effect magnitude rather than statistical significance.
  4. [§3.1–§3.2] The Pearson correlation analysis is performed on 'total energies' of interstitial configurations. If these are raw total energies from different SQS supercells, each of the three SQSs contributes a different defect-free reference energy, and correlations pooled across SQSs would mix host-lattice energies with interstitial formation energies. The manuscript does not state whether per-SQS reference energies were subtracted. Please clarify and, if needed, recompute the correlations using formation energies referenced to the defect-free SQS.
minor comments (6)
  1. [§2.1/§3.2] The definition of 'nearest-neighbor (NN) count of each host element (within 0.1 Å distance tolerance)' is ambiguous. Please specify whether the neighbor shell is defined relative to the minimum distance per configuration and justify the 0.1 Å tolerance.
  2. [§3.3] The six O configurations are described as 'randomly selected' but also as 'spanning the full energetic range predicted by Orb-v3.' A random draw would not generally span a range; please describe the actual selection protocol.
  3. [§1] Typo: 'user-friedly' should be 'user-friendly'.
  4. [§4.3] The phrase 'more than a factor of over 1400' is redundant. Use '>1400×' or 'more than a factor of 1400'.
  5. [§3.3/§4.3] The DFT benchmark uses the PBE functional, which is the same functional underlying MACE-MATPES-PBE-0. The agreement for O is therefore partly a consistency check; this should be acknowledged when describing the validation as independent.
  6. [§4.3] The manuscript describes the statistics as 'converged,' but no convergence analysis with respect to the number of SQSs or configurations is provided. A subset analysis would strengthen this claim.

Circularity Check

0 steps flagged

No significant circularity; the uMLIP results are inferences from external pretrained models checked against independent DFT calculations.

full rationale

The paper's derivation chain is: (1) build SQS supercells with ATAT (external, cited); (2) relax and score 6,750 interstitial configurations with three pretrained universal MLIPs (MACE-MATPES-PBE-0, Orb-v3, SevenNet-0), which are external models not fitted or fine-tuned in this work; (3) compute Pearson correlations between uMLIP energies and local chemical descriptors; and (4) benchmark six O interstitial configurations against DFT. No fitted parameter is renamed as a prediction, and no result is defined in terms of the quantity it claims to predict. The DFT benchmark is an independent check on the uMLIP energy ordering, even though MACE-MATPES-PBE-0 and DFT share the PBE functional, making that particular comparison a consistency check rather than a fully independent test of the PBE reference itself. The self-citations to SimplySQS and uMLIP-Interactive support tooling and reproducibility, not the load-bearing scientific claim of uMLIP transferability. The main non-circularity concern is one of external validity: the abstract asserts a DFT validation of H tetrahedral preference that is not presented in the main text (Section 3.3 only shows O benchmark), and no DFT checks are shown for C or N; this weakens the generalization of the conclusion but is not a circular reduction. Overall, the central energetic trends and site preferences are outputs of externally pretrained models, not consequences of the paper's own assumptions or fits, so the circularity score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The paper adds no new fitted parameters or invented physical entities; it relies on external uMLIPs and a small DFT validation set. The main hidden costs are the hand-chosen correlation thresholds and the assumption that the three SQSs and the uMLIPs faithfully represent the alloy and the interstitial energetics.

free parameters (3)
  • Pearson correlation significance threshold = ±0.1
    r values within -0.1 to 0.1 are treated as negligible in §3.2; this hand-chosen threshold determines which chemical trends are called significant.
  • Nearest-neighbor counting distance tolerance = 0.1 Å
    NN counts are defined within a 0.1 Å distance tolerance in §3.2; trends could shift if this tolerance changes.
  • SQS cluster cutoffs = pair ≤2.0 Å, triplet ≤1.5 Å
    Chosen in §2.1 to generate the three 250-atom SQS; cluster truncation affects how well the supercells represent random disorder.
axioms (4)
  • domain assumption uMLIPs pretrained on DFT data transfer to C/N/O/H interstitial energetics in Ti-23Nb-0.7Ta-2Zr without fine-tuning
    All 6,750 energies and relaxations come from the uMLIPs; only six O configurations are checked against DFT, so the accuracy of the potentials for the other elements and site types is assumed.
  • domain assumption Three independent 250-atom SQSs adequately represent the alloy's substitutional disorder
    The SQS cluster expansion is finite and only three 24-hour searches are used; no convergence test with respect to number or size of SQSs is reported.
  • domain assumption DFT-PBE is the reference ground truth for interstitial site preferences and energetic ordering
    DFT validation uses PBE VASP, and MACE-MATPES-PBE-0 is trained on PBE data, so the benchmark partly checks consistency within one functional rather than absolute accuracy.
  • domain assumption Known bcc site preferences (C/N/O octahedral, H tetrahedral) are the correct physical benchmark for judging SevenNet-0
    The paper labels SevenNet-0's octahedral H as a limitation based on expected bcc behavior and agreement between MACE and Orb; the abstract claims independent DFT H evidence, but it is not shown in the main text.

pith-pipeline@v1.3.0-alltime-deepseek · 10095 in / 9813 out tokens · 97049 ms · 2026-08-03T18:19:58.209955+00:00 · methodology

0 comments
read the original abstract

Understanding the behavior of light interstitial elements in multicomponent alloys remains challenging due to the complexity of local chemical environments and the high computational cost of first-principles calculations. Here we demonstrate that three universal machine-learning interatomic potentials (uMLIPs) - MACE-MATPES-PBE-0, Orb-v3, and SevenNet-0 can efficiently map the energetics of C, N, O, and H interstitials in a Ti-23Nb-0.7Ta-2Zr (at.%) gum metal base alloy while being several orders of magnitude faster than density functional theory (DFT). All uMLIPs predict broad energy distributions (~1-3 eV) across the four interstitial elements, reflecting their strong sensitivity to local lattice chemistry. Despite alloy disorder, MACE-MATPES-PBE-0 and Orb-v3 reproduce the expected site preferences of the bcc structure: C, N, and O relax into octahedral sites, whereas H stabilizes in tetrahedral positions. In contrast, SevenNet-0 predicts H to be most stable in octahedral coordination, indicating a limitation of this model. Correlation analysis reveals two dominant chemical trends: Ti-rich environments strongly stabilize interstitials, whereas close proximity to Nb is destabilizing. Zr and Ta show no statistically significant influence, likely due to their low concentrations. Benchmarking representative O interstitial configurations against DFT confirms that the uMLIPs reasonably reproduce the energetic ordering of chemically distinct environments. DFT validation confirms that tetrahedral configurations are energetically more favorable than octahedral sites for H interstitials, further illustrating the SevenNet-0 limitation. Overall, we demonstrated that uMLIPs enable computationally efficient, statistically broad characterization of defect energetics in Ti-23Nb-0.7Ta-2Zr gum metal base alloy and provide insight into how local chemical environments govern interstitial stability.

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

55 extracted references · 4 linked inside Pith

  1. [1]

    In bcc Ti-Nb-Zr based systems, light interstitials such as N, C, O, and H strongly influence phase stability and mechanical properties, even at low concentrations5–9

    Introduction Middle- and high -entropy alloys (M/HEAs) and other compositionally complex alloys typically exhibit exceptional combinations of strength, ductility, and thermal stability, largely due to the wide variations of local chemical environments present in their disordered atomic structures1–4. In bcc Ti-Nb-Zr based systems, light interstitials such...

  2. [2]

    The initial structure for the bcc phase was based on β-Ti (Im-3m, a = 3.282 Å35)

    Methods 2.1 SQS generation The SQS for Ti-23Nb-0.7Ta-2Zr gum metal base alloy were generated using the ATAT mcsqs module17,34. The initial structure for the bcc phase was based on β-Ti (Im-3m, a = 3.282 Å35). A 250-atom supercell of dimensions 5 × 5 × 5 was constructed, employing pair clusters up to 2.0 Å and triplet clusters up to 1. 5 Å (normalized rela...

  3. [3]

    Results 3.1 Energetics of interstitial configurations Evaluating all structurally-optimized 6750 interstitial configurations (750 octahedral and 1500 tetrahedral sites per each of three independent SQS, see also Fig. 1 for the employed workflow) with three uMLIPs (MACE-MATPES-PBE-0, Orb-v3, SevenNet-0) in Ti-23Nb-0.7Ta- 2Zr reveals a wide total energy ran...

  4. [4]

    C, N, and O consistently relax into octahedral sites, while H remains most stable in tetrahedral positions

    Discussion 4.1 Persistence of geometric site preferences The preference of C, N, and O for octahedral sites and of H for tetrahedral sites with MACE- MATPES-PBE-0 and Orb-v3 is in agreement with the typical behavior in bcc alloys. C, N, and O consistently relax into octahedral sites, while H remains most stable in tetrahedral positions. The persistence of...

    2025

  5. [5]

    Ti-rich environments are stabilizing All three uMLIPs show that increasing the number of Ti NN lowers the interstitial energy, as indicated by negative correlations (Fig. 4 ). MACE -MATPES-PBE-0 exhibits the strongest Ti sensitivity, particularly for H, followed by N, C, and O. Orb -v3 predicts the largest Ti correlation for O and N, with weaker trends fo...

  6. [6]

    Close Nb distances are destabilizing Nb exerts the opposite effect: shorter Nb interstitial distances correlate with higher energies, while increasing this distance lowers the energy (Fig. 5). Thus, the destabilizing role of Nb manifests primarily through distance rather than NN count. MACE-MATPES-PBE-0 and SevenNet-0 display similar trend ordering, with ...

  7. [7]

    Although the three uMLIPs exhibit consistent qualitative trends in how local chemical environments influence interstitial energetics, some quantitative differences remain

    Zr and Ta exert negligible influence Both Zr and Ta show no measurable correlation with interstitial energies. Although the three uMLIPs exhibit consistent qualitative trends in how local chemical environments influence interstitial energetics, some quantitative differences remain. MACE- MATPES-PBE-0 shows overall the strongest correlations between the ch...

  8. [8]

    Author contributions M.L. and J.D. conceived the original idea. M.L. wrote the initial draft, developed all codes, performed all calculations, and analyzed the results presented in the article. J.D. and P .V . supervised the project and contributed to the execution of the calcul ations, analysis of results, and improvement of the manuscript. V .M. improve...

  9. [9]

    Conclusions This work demonstrates that uMLIPs can reliably resolve the energetics of light interstitials (C, N, O, H) in the Ti -23Nb-0.7Ta-2Zr gum metal base alloy at a level of configurational sampling that is challenging and computationally demanding to reach with DFT . By evaluating total of 6750 interstitial configurations within three 250-atom SQSs...

  10. [10]

    TN02000069/003] and the Grant Agency of the Czech Technical University in Prague [grant No

    Acknowledgements This work was supported by the Technology Agency of the Czech Republic [grant No. TN02000069/003] and the Grant Agency of the Czech Technical University in Prague [grant No. SGS24/121/OHK2/3T/12]

  11. [11]

    Video tutorials demonstrating how to run calculations directly through its interface can be found here: youtu.be/xh98fQqKXaI?si=6KW1MqYNtPYEwiUR

    Data availability To ensure reproducibility of the results presented in this study, all scripts used to create the SQSs and to perform calculations with uMLIPs are publicly available in the following GitHub repository: github.com/bracerino/Ti-23Nb-0.7Ta-2Zr-Interstitials The tool used to perform calculations with uMLIPs, uMLIP-Interactive, is publicly acc...

  12. [12]

    Saito, T. et al. Multifunctional alloys obtained via a dislocation -free plastic deformation mechanism. Science (1979) 300, 464–467 (2003)

    1979

  13. [13]

    Competing interest The authors declare no competing interests

  14. [14]

    Miracle, D. B. & Senkov, O. N. A critical review of high entropy alloys and related concepts. Acta Mater 122, 448–511 (2017)

    2017

  15. [15]

    Ma, Y . et al. High-entropy energy materials: challenges and new opportunities. Energy Environ Sci 14, 2883–2905 (2021)

    2021

  16. [16]

    P ., Raabe, D

    George, E. P ., Raabe, D. & Ritchie, R. O. High-entropy alloys. Nat Rev Mater 4, 515–534 (2019)

    2019

  17. [17]

    N., Wilks, G

    Senkov, O. N., Wilks, G. B., Miracle, D. B., Chuang, C. P . & Liaw, P . K. Refractory high- entropy alloys. Intermetallics 18, 1758–1765 (2010)

    2010

  18. [18]

    Chen, Y . et al. Interstitial strengthening of refractory ZrTiHfNb 0. 5Ta0. 5Ox (x= 0.05, 0.1, 0.2) high-entropy alloys. Mater Lett 228, 145–147 (2018)

    2018

  19. [19]

    & Wang, H

    Jin, C., Li, X., Kang, J., Li, H. & Wang, H. Effect of interstitial oxygen/nitrogen on mechanical and wear properties of TiZrHfNb refractory high -entropy alloy. J Alloys Compd 960, 170863 (2023)

    2023

  20. [20]

    Zhang, C. et al. Effects of Interstitial Oxygen Content on Microstructures and Mechanical Properties of TiZrNb Refractory Medium -Entropy Alloy. Metals (2075 -

    2075

  21. [21]

    Wei, Q. et al. Influence of oxygen content on microstructure and mechanical properties of Ti–Nb–Ta–Zr alloy. Mater Des 32, 2934–2939 (2011)

    2011

  22. [22]

    Chong, Y . et al. Oxygen interstitials make metastable β titanium alloys strong and ductile. Acta Mater 257, 119165 (2023)

    2023

  23. [23]

    Kamimura, Y . et al. Thermally activated deformation of gum metal: A strong evidence for the peierls mechanism of deformation. Mater Trans 56, 2084–2087 (2015)

    2084

  24. [24]

    Nagasako, N. et al. Microscopic study of gum-metal alloys: A role of trace oxygen for dislocation-free deformation. Acta Mater 105, 347–354 (2016)

    2016

  25. [25]

    During the structural optimization, both the atomic positions and the cell (with fixed angles) were allowed to relax

    The force convergence criterion of 0.005 eV/ Å was employed. During the structural optimization, both the atomic positions and the cell (with fixed angles) were allowed to relax. To facilitate and standardize our simulations, we developed and used uMLIP -Interactive (github.com/bracerino/uMLIP-Interactive), an open -source graphical user interface (GUI) t...

  26. [26]

    Samin, A. J. A computational investigation of the interstitial oxidation thermodynamics of a Mo-Nb-Ta-W high entropy alloy beyond the dilute regime. J Appl Phys 128, (2020)

    2020

  27. [27]

    Song, Z. et al. First-principles studies of behavior of hydrogen and mechanical properties of TiCrTaV high-entropy alloys. Mater Today Commun 35, 105929 (2023)

    2023

  28. [28]

    Wu, C. et al. Hydrogen accommodation and its role in lattice symmetry in a TiNbZr medium-entropy alloy. Acta Mater 288, 120852 (2025)

    2025

  29. [29]

    Zunger, A., Wei, S.-H., Ferreira, L. G. & Bernard, J. E. Special quasirandom structures. Phys Rev Lett 65, 353 (1990)

    1990

  30. [30]

    van de Walle, A. et al. Efficient stochastic generation of special quasirandom structures. Calphad 42, 13–18 (2013)

    2013

  31. [31]

    Zhou, T. et al. Quantitative Mapping of Oxygen Affinity to Local Chemical Environments in Ti–Zr–Nb–Ta Alloys via Machine Learning. Inorg Chem (2025)

    2025

  32. [32]

    Deng, B. et al. CHGNet as a pretrained universal neural network potential for charge- informed atomistic modelling. Nat Mach Intell 5, 1031–1041 (2023)

    2023

  33. [33]

    & Ong, S

    Chen, C. & Ong, S. P . A universal graph deep learning interatomic potential for the periodic table. Nat Comput Sci 2, 718–728 (2022)

    2022

  34. [34]

    Batatia, I. et al. A foundation model for atomistic materials chemistry. J Chem Phys 163, (2025)

    2025

  35. [35]

    P ., Simm, G., Ortner, C

    Batatia, I., Kovacs, D. P ., Simm, G., Ortner, C. & Csányi, G. MACE: Higher order equivariant message passing neural networks for fast and accurate force fields. Adv Neural Inf Process Syst 35, 11423–11436 (2022)

    2022

  36. [36]

    Rhodes, B. et al. Orb-v3: atomistic simulation at scale. arXiv preprint arXiv:2504.06231 (2025)

    Pith/arXiv arXiv 2025

  37. [37]

    Yang, H. et al. Mattersim: A deep learning atomistic model across elements, temperatures and pressures. arXiv preprint arXiv:2405.04967 (2024)

    Pith/arXiv arXiv 2024

  38. [38]

    Kim, J. et al. Data-efficient multifidelity training for high -fidelity machine learning interatomic potentials. J Am Chem Soc 147, 1042–1054 (2024)

    2024

  39. [39]

    & Smidt, T

    Koker, T., Kotak, M. & Smidt, T. Training a Foundation Model for Materials on a Budget. arXiv preprint arXiv:2508.16067 (2025)

    arXiv 2025

  40. [40]

    Mazitov, A. et al. PET-MAD as a lightweight universal interatomic potential for advanced materials modeling. Nat Commun 16, 10653 (2025)

    2025

  41. [41]

    Kaplan, A. D. et al. A foundational potential energy surface dataset for materials. arXiv preprint arXiv:2503.04070 (2025)

    arXiv 2025

  42. [42]

    Barroso-Luque, L. et al. Open materials 2024 (omat24) inorganic materials dataset and models. arXiv preprint arXiv:2410.12771 (2024)

    Pith/arXiv arXiv 2024

  43. [43]

    & Marques, M

    Schmidt, J., Pettersson, L., Verdozzi, C., Botti, S. & Marques, M. A. L. Crystal graph attention networks for the prediction of stable materials. Sci Adv 7, eabi7948 (2021)

    2021

  44. [44]

    Horton, M. K. et al. Accelerated data -driven materials science with the Materials Project. Nat Mater 1–11 (2025)

    2025

  45. [45]

    C., Kaplan, A

    Kuner, M. C., Kaplan, A. D., Persson, K. A., Asta, M. & Chrzan, D. C. MP -ALOE: an r2SCAN dataset for universal machine learning interatomic potentials. NPJ Comput Mater 11, 352 (2025)

    2025

  46. [46]

    Jacobs, R. et al. A practical guide to machine learning interatomic potentials –Status and future. Curr Opin Solid State Mater Sci 35, 101214 (2025)

    2025

  47. [47]

    & Ceder, G

    van de Walle, A., Asta, M. & Ceder, G. The alloy theoretic automated toolkit: A user guide. Calphad 26, 539–553 (2002)

    2002

  48. [48]

    Levinger, B. W. Lattice parameter of beta titanium at room temperature. Jom 5, 195 (1953)

    1953

  49. [49]

    & van de Walle, A

    Lebeda, M., Drahokoupil, J., Vlčák, P ., Svoboda, Š. & van de Walle, A. Simplifying Generation of Special Quasirandom Structures with ATAT Using Interactive Online Interface-SimplySQS. arXiv preprint arXiv:2510.18020 (2025)

    Pith/arXiv arXiv 2025

  50. [50]

    & Furthmüller, J

    Kresse, G. & Furthmüller, J. Efficient iterative schemes for ab initio total -energy calculations using a plane-wave basis set. Phys Rev B 54, 11169 (1996)

    1996

  51. [51]

    P ., Burke, K

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

    1996

  52. [52]

    Larsen, A. H. et al. The atomic simulation environment —a Python library for working with atoms. Journal of Physics: Condensed Matter 29, 273002 (2017)

    2017

  53. [53]

    Casillas-Trujillo, L. et al. Interstitial carbon in bcc HfNbTiVZr high -entropy alloy from first principles. Phys Rev Mater 4, 123601 (2020)

    2020

  54. [54]

    & Jiao, Z

    Yu, Q., Qiu, S. & Jiao, Z. -B. Atomic-scale understanding of interstitial -strengthened high-entropy alloys: Q. Yu et al. Rare Metals 1–13 (2025)

    2025

  55. [55]

    & Song, X

    Chen, A., Pan, Y ., Dai, J., Fu, W. & Song, X. Theoretical study on the element distribution characteristics and the effects of oxygen in TiZrHfNb high entropy alloys. Mater Today Commun 36, 106922 (2023). Supplementary information The following supplementary figures present the energ etics of 6750 interstitial configurations for C, N, O, and H in Ti-23Nb...

    2023