Pith. sign in

REVIEW 3 major objections 4 minor 17 references

Local B-site chemistry controls oxygen-vacancy energetics in Ca-Ce-Ti-Mn perovskites for thermochemical hydrogen production

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Local B-site chemistry, not bulk stoichiometry, dominates oxygen-vacancy energetics in Ca–Ce–Ti–Mn perovskites for thermochemical hydrogen production.

desk verdict The B-site-dominates-Ev claim is solid and new; the Δδ screening payoff rests on an unverified low-Ev tail, so treat the ceria-beating numbers as provisional. read the letter →

arxiv 2607.28752 v1 pith:SRQJDBHI submitted 2026-07-30 cond-mat.mtrl-sci

Manish Kumar (1) , Natalia Ali (2) , Matthew D. Witman (3) , Shang Zhai (4) , James E. Miller (2) , Ivan Ermanoski (2) , Ellen B. Stechel (2) , Robert B. Wexler (1) ((1) Washington University in St. Louis
show 3 more authors
(2) Arizona State University (3) Sandia National Laboratories (4) The Ohio State University)
This is my paper · ORCID
classification cond-mat.mtrl-sci
keywords perovskiteoxygenvacancythermochemicalwatersplittinghydrogenproductionCCTMspecialquasirandomstructuredefectgraphneuralnetworkshort-rangecationorder
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

This paper tries to establish that, in Ca–Ce–Ti–Mn perovskite oxides, the energy needed to create an oxygen vacancy is set mainly by the two B-site metal cations directly bonded to that oxygen, not by the material's overall composition. Varying the local Mn fraction among nearest-neighbor B sites shifts the vacancy formation energy by 1.0–1.5 eV depending on local Ce content, whereas changing the nearby A-site Ce fraction shifts it by only 0.2–0.6 eV. If true, short-range ordering of B-site cations becomes a practical way to tune redox behavior without changing bulk stoichiometry. The authors support this with first-principles calculations that sample all fifteen distinct local oxygen environments, an interpretable two-descriptor model whose coefficients recover the bond-breaking and electron-stabilization picture, and thermogravimetric measurements on three compositions that track the predicted trend. A sympathetic reader would care because it suggests a concrete path toward lower-temperature solar hydrogen production: compositions near a Ce/Mn-balanced region are predicted to match or exceed ceria's per-cycle oxygen capacity at 1350 °C instead of roughly 1600 °C.

What carries the argument

The central object is the coverage-constrained special quasirandom structure: supercells constructed so that all fifteen symmetry-distinct oxygen first-nearest-neighbor environments—defined by the Ce fraction among four A-site neighbors and the Mn fraction among two B-site neighbors—are realized. The interpretable crystal-feature model then fits Ev to two local descriptors, the average metal–oxygen bond-dissociation energy and average crystal reduction potential over the six neighboring cations. Its fitted coefficients (approximately two bonds' worth of dissociation energy and stabilization of two released electrons) directly encode the thermochemical bond-breaking/reduction-stabilization pi

What would settle it

Compute Ev for two oxygen vacancies that have identical (xCe, xMn) first-nearest-neighbor environments but different second-shell Ti/Mn arrangements; if the energy difference exceeds about 0.2 eV, the single-shell model underpinning all reported distributions and Δδ values fails.

Watch

Extended reading notes

Core claim

The central claim is that oxygen-vacancy formation energy Ev in CCTM perovskites is controlled by the local first-nearest-neighbor cation environment—the identity of the two B-site cations (Ti vs Mn) and, secondarily, the four A-site cations (Ca vs Ce)—rather than by bulk stoichiometry alone. Increasing local Mn fraction lowers Ev by 1.0–1.5 eV depending on local Ce, while A-site Ce variation shifts Ev by only 0.2–0.6 eV. Because the B-site effect spans 2–3 times the width of the thermochemically favorable 3.4–3.9 eV window, the authors identify short-range B-site order as a candidate tuning lever. Composition-space maps then show that Ce/Mn-balanced compositions (XCe ≈ 0.29–0.33, XMn ≈ 0.58

Load-bearing premise

The whole screening pipeline assumes that the energy to create an oxygen vacancy depends only on its six immediately neighboring cations, with one representative vacancy standing in for each environment and ideal-mixing probabilities weighting them; if cations farther away also shift the energy, every reported distribution and cycle capacity is miscalibrated.

Editorial extensions

If this is right

  • Short-range B-site order, if kinetically retained, could narrow the Ev distribution and raise the fraction of sites inside the 3.4–3.9 eV target window at fixed bulk composition.
  • Composition-space screening identifies Ce/Mn-balanced compositions whose predicted per-cycle oxygen exchange at 1350 °C matches or exceeds ceria, which requires roughly 1600 °C.
  • Measured Δδ on three compositions increases monotonically with Ce content under protocols close to the model conditions, confirming the qualitative predicted trend and giving an experimental anchor for the screening.
  • Because the fitted model encodes bond-breaking and reduction-stabilization physics, the design rule transfers: dopants with weaker metal–oxygen bonds or more favorable reduction couples should lower Ev in related perovskite families.
  • The dominance of B-site chemistry implies that bulk stoichiometry screening alone is insufficient; local environment sampling is necessary to predict redox performance.
  • The find that B-site effects dominate implies bulk optimization averages over local environments, so targeted B-site ordering is a more effective lever than composition tuning alone.

Reading between the lines

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

  • If B-site ordering can be established, even compositions outside the identified Ce/Mn-balanced region could in principle be engineered to exceed the ideal-solution-predicted in-window fraction; this is testable by comparing ordered-motif DFT calculations with the SQS baseline.
  • The paper's worked example shows that using the mean Ev underestimates Δδ by an order of magnitude relative to the full environment distribution; by extension, the low-Ev tail of the site distribution governs performance, so synthesis and characterization should target tail properties, not just the average.
  • The disagreement between the linear crystal-feature model and the nonlinear graph-neural-network surface in mixed Ce–Mn regimes suggests some environment coupling beyond the first shell may exist; expanding the DFT dataset to intermediate XCe compositions would determine whether the ridge is physical or a sampling artifact.
  • If cerium is predominantly Ce3+ with electrons delocalized over many Ce 4f states rather than a discrete Ce4+→Ce3+ redox couple, then A-site Ce acts as an electron reservoir; this would change how the entropic and capacity benefits of Ce are modeled in other Ce-containing perovskites.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper combines SCAN+U DFT, a coverage-constrained special quasirandom structure (SQS) scheme, a linear crystal-feature model (CFM), a fine-tuned defect graph neural network (dGNN), and TGA/LSFR experiments to map oxygen-vacancy formation energies (Ev) and redox cycle capacities (Δδ) in Ca–Ce–Ti–Mn perovskites. The central mechanistic claim is that the local first-nearest-neighbor B-site Mn fraction controls Ev (1.0–1.5 eV shifts) more strongly than the local A-site Ce fraction (0.2–0.6 eV), with the six bulk compositions showing transferable local-environment trends. The screening layer identifies Ce/Mn-balanced compositions (XCe ≈0.29–0.33, XMn ≈0.58–0.67) whose predicted Δδ at 1350 °C matches or exceeds ceria, and three experimental compositions show Δδ increasing with XCe, qualitatively consistent with the models.

Significance. If the central mechanism claim holds, the paper provides a concrete design rule—short-range B-site cation order, not just bulk stoichiometry, can tune redox performance—and an interpretable framework transferable to other perovskite families. The coverage-constrained SQS methodology that realizes all fifteen 1NN environments is a useful technical contribution, and the public data release strengthens reproducibility. The CFM coefficients recovering a Born–Haber-like decomposition (βb ≈2.0, βr ≈−1.3) is a nice interpretability result. However, the quantitative screening claims (Fig. 5, the 'matches or exceeds ceria' benchmark in the abstract) rest on a 15-site DFT sample per composition with one vacancy per 1NN environment class, selected to be statistically typical. This sampling cannot determine the low-Ev tail that exponentially dominates the Boltzmann average underlying Δδ (Methods eqn (4)–(5); ESI S7). The paper's own numbers (dGNN 0.098 vs SCAN+U 0.044 vs measured 0.048–0.067 at CCTM2112; SCAN+U 0.069 vs measured 0.044–0.047 at XCe=0.29) show that tail errors appear in both directions and are composition-dependent. The central B-site-over-A-site hierarchy is plausible and survi

major comments (3)
  1. [Methods: Supercell construction and vacancy site selection; Methods eqn (4)–(5); ESI S7] The Δδ evaluations in Fig. 5 and Table 3 are exponentially sensitive to the low-Ev tail of each composition's Ev distribution, yet each of the fifteen 1NN environment classes is represented by exactly one vacancy chosen as the KDE mode of the second-shell composition. A mode-selected site is, by design, not informative about the tail. The paper's own cross-checks demonstrate this: at CCTM2112, the SCAN+U pipeline gives Δδ=0.044 while the dGNN, which scores every O site and sees beyond 1NN, gives 0.098; measured values are 0.048–0.067. At XCe=0.29 the SCAN+U pipeline overshoots (0.069 vs 0.044–0.047 measured) and the Discussion explicitly leaves open 'a nonrepresentative low-Ev tail in the 15-site DFT sample.' The tail is therefore unverified in both directions. Because the abstract's central performance claim ('matches or exceeds ceria at 1350 °C') and the Fig. 5 maps rely on these Δδ va
  2. [Model construction; ESI S3; Table 3] The CFM is a Huber regression fitted to all 90 DFT Ev values, so the 'predictions' in Fig. 5a and the CFM column of Table 3 are in-sample interpolations, not out-of-sample predictions. The leave-one-composition-out test in ESI S3 (MAE 0.174 eV) shows generalization to new bulk compositions under the assumed linear form, but it does not validate the form itself—in particular, it cannot test whether the 1NN-averaged descriptors with a composition-independent coefficient set capture the tail behavior that drives Δδ. Similarly, the dGNN was fine-tuned on 18 of the 90 CCTM sites, so its hold-out validation on the remaining 72 is better but still shares the same DFT reference. I recommend clearly labeling the CFM and dGNN results as fitted-model interpolations within the sampled composition window, and distinguishing the mechanistic interpretation of the CFM coefficients (which is robust) from
  3. [Fig. 4 and caption; Results: Local cation control of vacancy energetics] The claim that the local-environment dependence is 'largely transferable across bulk compositions' is based on the tight clustering of the six bulk markers in each (xCe, xMn) cell. But each marker is the Ev of a single vacancy chosen to be the KDE mode of the second-shell environment; there is no within-class, within-composition spread to assess. The clustering could therefore reflect the selection rule rather than physical transferability. This matters because the central hierarchy—B-site 1NN chemistry dominates Ev—is quantified from these single representatives, and the dGNN distributions (ESI Fig. S8) show continuous site-to-site variation beyond the 1NN description. Please provide a within-class variance estimate (e.g., from the dGNN's all-site predictions or from additional SCAN+U vacancies for a few classes) or explicitly state that the 1.0–1.5 eV and 0.2–0.6 eV ranges are single-s
minor comments (4)
  1. [Table 2] The column headings 'Ev ≤3.4 eV' and 'Ev ≤3.9 eV' are overlapping and not mutually exclusive; the 'In target range' column (3.4–3.9 eV) is the informative one. Consider renaming to 'Ev <3.4 eV' and '3.4–3.9 eV' or adding an explicit note that the first two columns are cumulative.
  2. [Methods: Supercell construction and vacancy site selection] The coverage penalty k=2000 in eqn (2) is introduced without a sensitivity check. Since the paper claims coverage is obtained at no cost in the SQS objective, please report the range of k for which all fifteen environments are realized, or at least state that the result is insensitive to k above some threshold.
  3. [Model construction: Defect graph neural network model; ESI S8.2] The fine-tuning dataset is described as a 'randomly drawn 20%' of the 90 CCTM sites. For exact reproducibility, specify the random seed and include the list of the 18 fine-tuning sites in the data release.
  4. [Results: Local cation control of vacancy energetics] The interpretation of the CFM coefficient βb ≈2.0 as 'approximately two bonds' worth' is slightly loose, because the descriptor ⟨Eb⟩ is averaged over all six nearest-neighbor cations; a single coefficient of 2.0 on the average does not uniquely identify two bonds. Consider rephrasing to 'the fitted coefficient corresponds to a total contribution of 2×⟨Eb⟩, which the authors associate with the two B-site M–O bonds.'

Circularity Check

2 steps flagged · score 4.0 of 10

In-sample fitted-model Δδ values inflate the screening layer, but B-site dominance rests on direct DFT and independent measurements.

  1. fitted input called prediction [Table 3 (CFM column); Methods 'Model construction — Crystal feature model'; Fig. 5a]
    "CFM: same binomial environment weighting applied to crystal-feature-model predictions of the per-environment Ev. ... The reported coefficients are obtained from a single Huber fit to the full 90-site dataset."

    The CFM's per-environment Ev values are the outputs of a Huber regression trained on all 90 SCAN+U Ev labels at these same environments. The Table 3 CFM Δδ is the Boltzmann average (Methods eqn (4)-(5)) of those fitted values, so it is an in-sample transformation of the training data, not an out-of-sample prediction. The Discussion likewise describes the CFM as operating 'within its training composition window.' LOGO cross-validation is reported only as an error estimate; the screening surface and Table 3 use the full-data fit, so the 'predicted cycle capacity matches or exceeds ceria' claim inherits the in-sample character.

  2. fitted input called prediction [Methods 'Defect graph neural network model'; ESI S8.2/S8.4; Fig. 5b; Table 3]
    "we constructed a fine-tuning (FT) dataset by augmenting the Base data with 20% of the CCTM vacancy energies computed in this work. ... Using K-fold cross-validation, each Ev is predicted by an ensemble of K models; the ensemble mean provides the final prediction."

    The dGNN Δδ values are labeled 'predictions,' but for CCTM sites drawn into the 18-site fine-tuning set, the ensemble mean averages K models of which K−1 were trained on that site (only the model with the site in its held-out fold has not seen it). Thus the ensemble-mean dGNN values in Fig. 5b and Table 3, including the 0.098 outlier at CCTM2112, are not leave-one-out predictions; they are partly in-sample fits. This is less central than the CFM issue because the dGNN is a structural model and most CCTM sites are held out, but presenting the ensemble mean as 'prediction' overstates its independence.

full rationale

The paper's central mechanistic result — that nearest-neighbor B-site Mn fraction shifts Ev by 1.0–1.5 eV while A-site Ce shifts it by only 0.2–0.6 eV — is derived directly from SCAN+U DFT vacancy energies grouped by local environment (Fig. 4). That claim does not reduce to the fitted CFM or dGNN; it is an independent first-principles observation and is not circular. The experimental TGA/LSFR measurements were not used to fit either model and independently show the monotonic Ce trend, anchoring the qualitative conclusion. The circularity concerns are confined to the screening layer. The CFM column in Table 3 and the CFM-based Δδ maps are in-sample outputs of a regression fitted to all 90 DFT Ev values, so presenting them as 'predictions' of Δδ at those same compositions is a fitted-input-called-prediction step. The dGNN screening surface is also partly in-sample because the reported ensemble mean for fine-tuned CCTM sites includes models that trained on those very sites. The paper's own Discussion flags the low-Ev tail and nonrepresentative sampling ('a nonrepresentative low-Ev tail in the 15-site DFT sample'), but that is an accuracy limitation, not a circularity. The referenced 'target 3.4–3.9 eV window established in our earlier thermodynamic analysis' is a self-citation, but it is not load-bearing for the Δδ predictions, which use explicit T and PO2 conditions. Overall, the mechanism claim survives, while the quantitative screening claims are partially in-sample fits; hence score 4 rather than 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim is supported by a large DFT dataset (90 vacancy energies) and two fitted models. The main free parameters are the CFM coefficients, dGNN weights, and the hand-chosen SQS penalty. The structural axioms are the 1NN-completeness assumption, the neutral-noninteracting-vacancy model, the binomial cation statistics, and the transferability of SCAN+U with fixed U values.

free parameters (3)
  • CFM regression coefficients (βb, βr, β0) = 2.00, -1.34, -4.26 eV (SCAN+U feature parameterization)
    Huber fit to the 90 DFT oxygen-vacancy formation energies (ESI S3); used to compute Ev and Δδ across composition space, including at the six DFT compositions (in-sample).
  • dGNN model weights = trained on 1859 vacancy energies (Base) + 18 CCTM sites
    Neural-network parameters fitted to the augmented dataset; used for the screening surface and Table 3 predictions.
  • SQS coverage penalty k = 2000
    Hand-chosen factor in the modified SQS objective (Methods eqn 2) to force complete 1NN coverage; no sensitivity analysis reported.
assumptions (5)
  • domain assumption SCAN+U with UCe=2.0, UTi=2.5, UMn=2.7 eV gives accurate oxygen-vacancy formation energies for CCTM.
    U values taken from prior benchmarks (refs 56–57); the paper does not sweep U for CCTM and notes magnetic/chemical disorder precludes exhaustive checks.
  • domain assumption The first-nearest-neighbor cation environment fully determines the vacancy formation energy distribution; second-shell and longer-range effects are negligible.
    Methods: one representative vacancy per 1NN class; ESI eqn S13–S14 weights sites by binomial 1NN probabilities only. This is the load-bearing completeness assumption of the screening pipeline.
  • domain assumption Vacancies are neutral, noninteracting, with temperature-independent Ev; the host is metallic.
    Methods eqn (4)–(5); ESI S4 bounds vacancy–vacancy electrostatics at few-tens-of-meV scale. The metallic-host assumption is supported by DOS (ESI Fig. S2).
  • domain assumption Ideal-solution binomial statistics describe the bulk cation distribution in the SQS supercells.
    ESI eqn (S13); the SQSs are generated to approximate random disorder, but short-range order could deviate from binomial.
  • standard math Standard lattice-gas statistical mechanics and NIST-JANAF thermochemistry apply.
    Methods eqns (4)–(8) and ESI S7; standard model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Local B-site chemistry controls oxygen-vacancy energetics in Ca-Ce-Ti-Mn perovskites for thermochemical hydrogen production." pith.science (2026). https://pith.science/paper/SRQJDBHI

@misc{pith2026260728752,
  author       = {Pith},
  title        = {Pith review of: Local B-site chemistry controls oxygen-vacancy energetics in Ca-Ce-Ti-Mn perovskites for thermochemical hydrogen production},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SRQJDBHI}},
  note         = {Machine review of arXiv:2607.28752}
}
abstract

Two-step thermochemical water splitting driven by concentrated solar heat is a scalable route to renewable hydrogen, but it requires oxides whose oxygen-vacancy formation energies balance facile reduction with favorable reoxidation. Perovskite solid solutions can tune this balance, but the relationship between bulk stoichiometry and local defect energetics remains poorly understood. Here we map oxygen-vacancy formation energetics across Ca-Ce-Ti-Mn (CCTM) perovskites by combining first-principles calculations with a coverage-constrained special quasirandom structure approach that realizes all fifteen symmetry-distinct oxygen nearest-neighbor environments, an interpretable crystal-feature model whose fitted coefficients directly encode the underlying Born-Haber thermochemistry, and a fine-tuned defect graph neural network. Local B-site chemistry dominates the oxygen-vacancy formation energy $E_\mathrm{v}$: varying the nearest-neighbor Mn fraction shifts $E_\mathrm{v}$ by 1.0-1.5 eV depending on local Ce content, whereas A-site Ce variation contributes a smaller, Mn-dependent shift of 0.2-0.6 eV. Short-range B-site cation order, if it can be established and kinetically retained through processing, is therefore a candidate means of tuning redox performance without changing bulk composition. Composition-space maps identify a Ce/Mn-balanced region ($X_\mathrm{Ce}$ = 0.29-0.33, $X_\mathrm{Mn}$ = 0.58-0.67) combining a high fraction of vacancy sites within the targeted $E_\mathrm{v}$ window with phase stability and solubility, whose predicted redox cycle capacity matches or exceeds the ceria benchmark at 1350 $^\circ$C rather than the roughly 1600 $^\circ$C ceria requires. Measurements on three CCTM compositions show cycle capacity increasing monotonically with Ce content under protocols close to the model conditions. The design rules are expected to transfer to related perovskite families.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 1 canonical work pages

  1. [1]

    R. B. Wexler, G. Sai Gautam, R. T. Bell, S. Shulda, N. A. Strange, J. A. Trindell, J. D. Sugar, E. Nygren, S. Sainio, A. H. McDaniel, D. Ginley, E. A. Carter and E. B. Stechel,Energy & Environmental Science, 2023,16, 2550–2560

    2023

  2. [2]

    B. Xu, J. Park, D. Zhang, H. A. De Santiago, W. Li, X. Liu, J. Luo, S. Lany and Y. Qi, Chemistry of Materials, 2024,36, 4990–5001

    2024

  3. [3]

    Sai Gautam and E

    G. Sai Gautam and E. A. Carter,Physical Review Materials, 2018,2, 095401

    2018

  4. [4]

    Henkelman, A

    G. Henkelman, A. Arnaldsson and H. J´ onsson,Computational Materials Science, 2006,36, 354–360

    2006

  5. [5]

    W. Tang, E. Sanville and G. Henkelman,Journal of Physics: Condensed Matter, 2009,21, 084204

    2009

  6. [6]

    R. B. Wexler, G. S. Gautam, E. B. Stechel and E. A. Carter,Journal of the American Chemical Society, 2021,143, 13212–13227

    2021

  7. [7]

    M. D. Witman and P. Schindler,Digital Discovery, 2025,4, 625–635

    2025

  8. [8]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot and ´E. Duchesnay,J. Mach. Learn. Res., 2011,12, 2825–2830

    2011

Show all 17 references
  1. [9]

    S. P. Ong, W. D. Richards, A. Jain, G. Hautier, M. Kocher, S. Cholia, D. Gunter, V. L. Chevrier, K. A. Persson and G. Ceder,Computational Materials Science, 2013,68, 314–319

    2013

  2. [10]

    Baldassarri, J

    B. Baldassarri, J. He and C. Wolverton,Physical Review Materials, 2024,8, 055407

    2024

  3. [11]

    Lany,Journal of the American Chemical Society, 2024, jacs.4c02688

    S. Lany,Journal of the American Chemical Society, 2024, jacs.4c02688

    2024

  4. [12]

    J. Park, B. Xu, J. Pan, D. Zhang, S. Lany, X. Liu, J. Luo and Y. Qi,npj Computational Materials, 2023,9, 29. 18 [13]NIST Chemistry WebBook, NIST Standard Reference Database Number 69, 2025,https: //webbook.nist.gov/chemistry/, Online; accessed 2025

    2023

  5. [14]

    Chase, Malcolm W.,NIST-JANAF Thermochemical Tables, American Institute of Physics, Woodbury, NY, 4th edn, 1998

    J. Chase, Malcolm W.,NIST-JANAF Thermochemical Tables, American Institute of Physics, Woodbury, NY, 4th edn, 1998

    1998

  6. [15]

    M. D. Witman, A. Goyal, T. Ogitsu, A. H. McDaniel and S. Lany,Nature Computational Science, 2023,3, 675–686

    2023

  7. [16]

    Xie and J

    T. Xie and J. C. Grossman,Physical Review Letters, 2018,120, 145301

    2018

  8. [17]

    T. C. Douglas, M. J. Dzara, A. J. E. Rowberg, K. A. King, M. Syrigou, N. A. Strange, R. T. Bell, A. Goyal, P.-W. Guan, R. B. Wexler, J. B. Varley, T. Ogitsu, S. Lany, A. H. McDaniel, S. R. Bishop and M. D. Witman,Materials Horizons, 2026,13, 829–839

    2026

  9. [18]

    M. D. Witman, S. Pujet, A. J. E. Rowberg, C. Sutton, J. B. Varley, S. Lany and R. B. Wexler, Transfer learning on universal interatomic potential embeddings improves generalization in structure-property defect models, 2026,https://doi.org/10.26434/chemrxiv.15003938/v1, Preprin...

    2026 doi

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.