Pith. sign in

REVIEW 3 major objections 6 minor 42 references

BatteryMat: a hierarchical machine-learning and DFT framework for average-voltage screening of lithium-ion cathode materials

T0 review · 3 major / 6 minor · reviewed 2026-07-11 · grok-4.5

Pith's one-line read A three-tier machine-learning and DFT pipeline ranks lithium-ion cathode candidates by average voltage and recovers experimental voltages within 0.3 V on commercial chemistries.

desk verdict Solid, carefully scoped hierarchical cathode screen: DFT tier matches experiment on four commercial chemistries; the 0.17 V figure is distillation of ALIGNN-FF, not DFT, and the shortlists are openly surrogate leads. read the letter →

arxiv 2607.06645 v1 pith:WKPA7EVI submitted 2026-07-07 cond-mat.mtrl-sci

classification cond-mat.mtrl-sci
keywords lithium-ioncathodeshigh-throughputscreeninggraphneuralnetworksALIGNNDFTaveragevoltagedelithiationJARVIS
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

Average intercalation voltage is the primary figure of merit for cathode screening, yet full DFT delithiation curves cannot scale to modern materials databases, while pure machine-learning surrogates skip thermodynamic consistency. BatteryMat places a single-pass graph-network voltage predictor at the front of a three-tier cascade, then validates survivors with force-field delithiation profiles and automated supercell DFT that chooses the exchange-correlation functional by spacegroup and recomputes the lithium-metal reference in the same plane-wave basis. On four stoichiometric commercial cathodes the DFT tier matches experimental average voltage to within 0.3 V and theoretical volumetric capacity to within 5 percent. Applied to existing databases the pipeline returns 71 lithium candidates from JARVIS-DFT and 213 from roughly 4.49 million Alexandria structures—all prioritised surrogate-level leads rather than confirmed new materials. The practical claim is that average voltage can drive high-throughput ranking without sacrificing experimental agreement at the validation end.

What carries the argument

The three-tier hierarchy: an ALIGNN scalar average-voltage predictor that collapses an N-step force-field protocol into one forward pass (0.17 V MAE against its ALIGNN-FF labels), an ALIGNN-FF step-by-step delithiation tier that ranks lithium vacancies, and automated supercell DFT with automatic functional selection and a recomputed lithium chemical potential.

What would settle it

Run the full DFT tier on the top-ranked non-commercial shortlist entries; if their computed average voltages deviate systematically by more than about 0.5 V from independent high-quality experiment or calculation while the four commercial benchmarks remain accurate, the surrogate primary filter has failed.

Watch

Extended reading notes

Core claim

BatteryMat establishes that a single-pass ALIGNN regressor, trained on 7,610 force-field delithiation voltages, can serve as the primary screening signal for lithium-ion cathodes; when survivors are advanced through ALIGNN-FF profiles and automated PBE+U or optB88-vdW+U supercell DFT (with spacegroup-selected functionals and an in-house lithium reference that removes a roughly 1 V tabulated offset), the pipeline recovers experimental average voltages to within 0.3 V and crystallographic volumetric capacities to within 5 percent on four commercial chemistries, while ranking existing database structures into shortlists of prioritised candidates rather than inventing new ones.

Load-bearing premise

The cascade assumes that the force-field delithiation protocol ranks lithium vacancies and plateaus faithfully enough that a single neural-network pass distilled from those labels still orders real cathode candidates usefully.

Editorial extensions

If this is right

  • Database-scale cathode ranking becomes feasible in minutes on a single GPU before any DFT budget is spent.
  • Layered frameworks are automatically routed to optB88-vdW+U and 3D-bonded frameworks to PBE+U without per-material manual choice.
  • Recomputing the lithium-metal reference in the cathode plane-wave basis removes a systematic ~1 V offset and places voltages on a common experimental scale.
  • The 71 JARVIS and 213 Alexandria shortlists become the concrete input for the next round of DFT validation campaigns.
  • For the four chemistries with DFT residuals below 0.3 V, cell-level energy density estimated as V_avg times theoretical capacity stays within roughly 8 percent of experiment.

Reading between the lines

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

  • Ensembling several universal force fields at the second tier would turn vacancy-ranking disagreement into an explicit confidence band before DFT is spent.
  • Retraining the scalar head on true DFT voltages instead of force-field labels would convert the 0.17 V figure into a direct DFT-distillation error and change how users interpret the primary screen.
  • Once a few sodium and magnesium DFT anchors exist, the same hierarchy can be re-ranked for multivalent hosts, because the surrogate already recovers the expected voltage–capacity trade-off for those ions.
  • Finite-supercell staircases on two-phase materials such as LiFePO4 imply that larger cells or explicit phase-boundary models will be required before the method can claim plateau-shape fidelity rather than only average-voltage fidelity.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. BatteryMat is a three-tier hierarchical pipeline for average-voltage screening of Li-ion cathodes: (i) a single-pass ALIGNN scalar regressor trained on 7,610 ALIGNN-FF delithiation labels (MAE 0.17 V, R²=0.94 vs those labels), (ii) ALIGNN-FF step-by-step delithiation profiles, and (iii) automated supercell DFT with spacegroup-selected PBE+U or optB88-vdW+U and an in-house Li-metal reference that removes a ~1 V tabulated offset. On four stoichiometric commercial cathodes the DFT tier recovers experimental average voltage to within 0.3 V and theoretical volumetric capacity to within 5%; a non-stoichiometric NMC-like entry is treated as an edge case. The pipeline prioritises existing structures, returning 71 JARVIS-DFT and 213 Alexandria surrogate-level leads rather than confirmed new cathodes. Code and a web demo are released.

Significance. If the hierarchy is reliable, the work is a useful, reproducible methods contribution: it couples a carefully scoped ML front-end to a transparent DFT validation tier, with practical engineering choices (auto functional selection, recomputed μ_Li, inherited block-AFM MAGMOM) that other high-throughput cathode campaigns can adopt. Strengths include explicit scoping of the 0.17 V figure as distillation fidelity rather than DFT agreement, open code and AtomGPT web access, and a clean four-chemistry experimental benchmark. The paper does not claim generative discovery; its value is prioritisation at database scale with a documented validation path.

major comments (3)
  1. [Sec. 2.1–2.2, 2.6, 2.9; Table 2] Sec. 2.1–2.2, 2.6, 2.9 and Table 2: The primary screening signal is a distillation of ALIGNN-FF labels. Sec. 2.6 already shows only group-level (not rank-level) agreement with DFT and within-group order swaps (LFP/LMP, LMO/LCO). The top-12 shortlist is dominated by polyanion phosphates and fluorides that are outside the five DFT benchmarks. Without at least a few DFT (or ALIGNN-FF vs DFT step-voltage) checks on those host families, the claim that single-pass ALIGNN is a useful primary ranking signal for the reported shortlists remains under-supported, even though the paper correctly labels them as surrogate-level leads.
  2. [Sec. 2.8] Sec. 2.8: Four of the five commercial benchmarks are removed by the Q_grav,min = 20 mAh/g filter because JARVIS reduced-cell capacities are ~9–11 mAh/g; only LCO survives, at rank 37/71. A screen that systematically excludes known working cathodes under its default thresholds needs either a formula-unit renormalisation, a revised default threshold with a sensitivity analysis, or a clear demonstration that the composite score still recovers known chemistries when the capacity metric is put on a conventional scale. As written, this weakens confidence that the ranking protocol would surface useful candidates in practice.
  3. [Sec. 2.4; Sec. 2.1] Sec. 2.4 vs Sec. 2.1: The Alexandria funnel derives surrogate voltages from ALIGNN formation-energy differences (Materials Project battery-explorer style) after ALIGNN-FF relaxation of charged hosts, whereas the JARVIS ALIGNN head is trained on sequential ALIGNN-FF vacancy-removal averages. These are different protocols. The manuscript should state explicitly whether the two shortlists are comparable, and whether the 0.17 V distillation metric applies to the Alexandria voltages at all (it does not, by construction).
minor comments (6)
  1. [Abstract; Sec. 2.1] Abstract and Sec. 2.1: The phrasing ‘promotes single-pass average-voltage prediction with ALIGNN as the primary screening signal’ is easy to misread as ALIGNN-vs-experiment accuracy. Consider leading every abstract/results mention of 0.17 V with ‘vs ALIGNN-FF labels’ in the same sentence.
  2. [Table 1; Sec. 2.5.4] Table 1 / Sec. 2.5.4: NMC was run with PBE+U although the auto-selector would now route R-3m to optB88-vdW+U. Flagging this as future work is fine, but the table caption should state the functional mismatch more prominently so readers do not treat the +0.70 V residual as a pure framework failure.
  3. [Fig. 6; Sec. 2.6] Fig. 6 and Sec. 2.6: ALIGNN-FF volumetric capacities use unrelaxed JARVIS volumes while DFT uses relaxed CONTCAR volumes. A one-sentence note in the figure caption would prevent over-interpreting the capacity panel as a pure method comparison.
  4. [Sec. 2.1] Sec. 2.1: ‘None of the five DFT-validated benchmark cathodes appear in the held-out test partition’ is good; also state that four fall in train and one in validation so readers do not assume full leave-out of commercial chemistries from the entire training process.
  5. [Sec. 5] Methods, lithium reference: Report the BCC Li lattice constant and k-mesh density used for μ_Li so the in-house values (−1.9031 / −0.9646 eV) are fully reproducible without reverse-engineering.
  6. [Throughout] Typos / notation: ‘V oltage’ with a space appears in several places (e.g. Introduction); ‘JV ASP’ inconsistently spaced; Eq. (1) uses μ_Li while the text sometimes writes μ Li. Standardise.

Circularity Check

1 steps flagged · score 1.0 of 10

No load-bearing circularity: the 0.17 V figure is an explicitly disclosed distillation metric, and DFT-vs-experiment claims are independent of the surrogate labels.

  1. fitted input called prediction [Abstract; Sec. 2.1 / 2.1.2 (label generation and parity)]
    "Trained on 7,610 ALIGNN-FF delithiation voltages, the ALIGNN predictor reproduces the force-field labels with a mean absolute error of 0.17 V and a coefficient of determination of 0.94; this measures distillation fidelity to the force-field protocol, not agreement with DFT or experiment. ... Because the labels are produced by ALIGNN-FF rather than DFT, the parity error of the ALIGNN scalar regressor measures how faithfully a single forward pass reproduces the multi-step force-field protocol; it does not measure ALIGNN-vs-DFT error directly."

    The 0.17 V / R²=0.94 headline is agreement of a model with labels generated by the same ALIGNN-FF protocol it distills; on the held-out split this is ordinary supervised test error, not an independent voltage prediction. The paper discloses the reduction and does not use it as experimental validation, so the step is minor and non-load-bearing for the DFT-vs-experiment claims.

full rationale

BatteryMat's derivation chain does not force any external claim by construction. The ALIGNN scalar head is trained on in-house ALIGNN-FF delithiation averages and evaluated against a held-out split of those same labels (MAE 0.17 V, R²=0.94). That number is therefore a standard train/test distillation fidelity score, not a prediction of DFT or experiment; the abstract and Sec. 2.1.2 state this explicitly and point the honest external test to the five-cathode DFT benchmark. The DFT tier (PBE+U / optB88-vdW+U supercell delithiation with an in-house Li reference) is compared to independent experimental voltages and crystallographic capacities; nothing in the training labels or the ALIGNN fit enters those total energies. The 71- and 213-candidate shortlists are labeled as surrogate-level prioritizations awaiting DFT, not confirmed cathodes. Self-citations (ALIGNN, JARVIS, AtomGPT) supply tools and databases, not uniqueness theorems or load-bearing premises that forbid alternatives. The softest condition in the paper—whether ALIGNN-FF vacancy ranking tracks DFT outside the five benchmarks—is a correctness/transferability risk, not a circular reduction. Score 1 only to mark the minor, fully disclosed self-consistency character of the lead MAE; the central experimental claims stand on independent footing.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claims rest on standard DFT+U practice, Materials-Project Hubbard U values, the assumption that ALIGNN-FF vacancy energetics are rank-preserving for screening, and a handful of hand-chosen screening thresholds. No new physical entities are postulated; free parameters are either taken from prior literature or are explicit filter cut-offs.

free parameters (4)
  • Materials Project Hubbard U_eff values (Mn 3.90, Fe 5.30, Co 3.32, Ni 6.20 eV, …)
    Taken unchanged from Jain et al.; residuals of opposite sign on LFP versus LMP are attributed to this fixed calibration.
  • Screening windows (V_avg 3.0–4.5 V, Q_grav > 20 mAh/g, E_hull < 0.05 eV/atom, V_max < 5.5 V)
    Hand-chosen thresholds that determine which of the 7 474 lithium entries survive into the 71-candidate list.
  • Composite score weights (1/3 V_avg + 1/3 Q_grav – 1/3 E_hull after min–max normalisation)
    Arbitrary equal weighting that ranks the final shortlist; changing the weights reorders candidates.
  • Supercell minimum lattice length 7 Å and 300-atom soft cap
    Engineering choice that fixes the 2 imes2 imes1 / 2 imes2 imes2 cells used for all five benchmarks.
assumptions (4)
  • domain assumption PBE+U (or optB88-vdW+U) total-energy differences yield average intercalation voltages accurate to ~0.1–0.3 V when the Li reference is computed in the same basis.
    Standard high-throughput battery DFT premise (Aydinol, Ceder, Hautier et al.); invoked throughout Sec. 2.5–2.6 and Methods.
  • ad hoc to paper Spacegroup membership (R-3m, P6_3/mmc, C2/m, C2/c versus 3-D bonded) is a reliable automatic selector between optB88-vdW+U and PBE+U.
    Introduced in Sec. 2.5 / Methods; NMC was run before the rule existed and is flagged as a caveat.
  • domain assumption ALIGNN-FF single-point vacancy energies correctly rank the lowest-energy lithium site at each delithiation step for screening purposes.
    Load-bearing for both label generation (Sec. 2.1.2) and the second-tier profiles (Sec. 2.2).
  • ad hoc to paper Block-AFM MAGMOM patterns inherited across steps preserve the correct magnetic sublattice and avoid SCF divergence.
    Implementation detail discovered during the campaign (Methods); not previously standard in the public screening literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BatteryMat: a hierarchical machine-learning and DFT framework for average-voltage screening of lithium-ion cathode materials." pith.science (2026). https://pith.science/paper/WKPA7EVI

@misc{pith2026260706645,
  author       = {Pith},
  title        = {Pith review of: BatteryMat: a hierarchical machine-learning and DFT framework for average-voltage screening of lithium-ion cathode materials},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WKPA7EVI}},
  note         = {Machine review of arXiv:2607.06645}
}
read the original abstract

Density functional theory (DFT) predicts cathode voltages accurately but does not scale to the combinatorial chemical spaces of modern materials databases, while pure machine-learning surrogates are fast but cannot guarantee thermodynamic consistency. We introduce BatteryMat, a three-tier framework that promotes single-pass average-voltage prediction with the Atomistic Line Graph Neural Network (ALIGNN) as the primary screening signal across JARVIS-DFT, then validates survivors with ALIGNN-FF force-field delithiation profiles and automated PBE+U or optB88-vdW+U supercell DFT. The exchange-correlation functional is selected automatically by spacegroup, and the lithium metal reference is recomputed in the same plane-wave basis as the cathode runs, removing a systematic offset of about 1 V present in tabulated values. Trained on 7,610 ALIGNN-FF delithiation voltages, the ALIGNN predictor reproduces the force-field labels with a mean absolute error of 0.17 V and a coefficient of determination of 0.94; this measures distillation fidelity to the force-field protocol, not agreement with DFT or experiment. On four commercial chemistries (LiFePO4, LiMnPO4, LiMn2O4, LiCoO2) the DFT tier reproduces the experimental average voltage to within 0.3 V and the theoretical volumetric capacity to within 5%; a fifth, non-stoichiometric layered entry is carried as an edge case. The pipeline prioritises, rather than generates, existing structures: it ranks the lithium-containing JARVIS-DFT pool into 71 candidates and a scan of about 4.49 million Alexandria structures into 213, all surrogate-level leads awaiting DFT validation rather than confirmed cathodes. BatteryMat is available at https://github.com/atomgptlab/batterymat with a demo at https://atomgpt.org/battery.

Figures

Figures reproduced from arXiv: 2607.06645 by the authors.

Figure 1
Figure 1. Hierarchical schematic of the BatteryMat pipeline. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. ALIGNN average-voltage parity plot. Predicted versus target average voltage on the held-out test set after 250 training epochs. The dashed line is y = x; the dotted line is the dataset-mean baseline. MAE = 0.17 V; R 2 = 0.94. hyperparameters on 7,610 lithium-containing entries from JARVIS-DFT. The training labels are not raw DFT voltages: JARVIS-DFT does not contain a per-material voltage subset, and the average vol… view at source ↗
Figure 3
Figure 3. Foundation-scale cathode screening funnel over Alexandria PBE 3D. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Voltage–capacity landscape of the 213-candidate Alexandria shortlist. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: ALIGNN-FF versus DFT for LiCoO2 as the prototype layered cathode. Left: ALIGNN-FF predicted voltage profile. Right: DFT voltage profile for LixCo8O16 with three staged plateaus at 4.01/4.23/4.48 V, consistent with the multi-step staging observed experimentally37 [PIT…
Figure 6
Figure 6. Figure 6: Voltage and volumetric-capacity benchmark. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 42 canonical work pages

  1. [1]

    Whittingham, M. S. Electrical energy storage and intercalation chemistry.Science192, 1126–1127 (1976)

  2. [2]

    C., Wiseman, P

    Mizushima, K., Jones, P. C., Wiseman, P. J. & Goodenough, J. B. Li xCoO2 (0<x≤1 ): A new cathode material for batteries of high energy density.Mater . Res. Bull.15, 783–789 (1980). 27

  3. [3]

    K., Nanjundaswamy, K

    Padhi, A. K., Nanjundaswamy, K. S. & Goodenough, J. B. Phospho-olivines as positive-electrode materials for rechargeable lithium batteries.J. Electrochem. Soc.144, 1188–1194 (1997)

  4. [4]

    & Makimura, Y

    Ohzuku, T. & Makimura, Y . Layered lithium insertion material of LiCo 1/3Ni1/3Mn1/3O2 for lithium-ion batteries.Chem. Lett.30, 642–643 (2001)

  5. [5]

    The birth of the lithium-ion battery.Angew

    Yoshino, A. The birth of the lithium-ion battery.Angew. Chem. Int. Ed.51, 5798–5800 (2012)

  6. [6]

    Goodenough, J. B. & Park, K.-S. The Li-ion rechargeable battery: A perspective.J. Am. Chem. Soc. 135, 1167–1176 (2013)

  7. [7]

    K., Kohan, A

    Aydinol, M. K., Kohan, A. F., Ceder, G., Cho, K. & Joannopoulos, J. Ab initio study of lithium intercalation in metal oxides and metal dichalcogenides.Phys. Rev. B56, 1354–1365 (1997)

  8. [8]

    Jain, A.et al.Formation enthalpies by mixing GGA and GGA+ u calculations.Phys. Rev. B84, 045115 (2011)

Show all 42 references
  1. [9]

    Mater .23, 3495–3508 (2011)

    Hautier, G.et al.Phosphates as lithium-ion battery cathodes: An evaluation based on high- throughputab Initiocalculations.Chem. Mater .23, 3495–3508 (2011)

  2. [10]

    Jain, A.et al.Commentary: The Materials Project: A materials genome approach to accelerating materials innovation.APL Mater .1, 011002 (2013)

  3. [11]

    Curtarolo, S.et al.AFLOW: An automatic framework for high-throughput materials discovery. Comput. Mater . Sci.58, 218–226 (2012)

  4. [12]

    E., Kirklin, S., Aykol, M., Meredig, B

    Saal, J. E., Kirklin, S., Aykol, M., Meredig, B. & Wolverton, C. Materials design and discovery with high-throughput density functional theory: The Open Quantum Materials Database (OQMD). JOM65, 1501–1509 (2013)

  5. [13]

    Mater .6, 173 (2020)

    Choudhary, K.et al.The joint automated repository for various integrated simulations (JARVIS) for data-driven materials design.npj Comput. Mater .6, 173 (2020)

  6. [14]

    The jarvis infrastructure is all you need for materials design.Comput

    Choudhary, K. The jarvis infrastructure is all you need for materials design.Comput. Mater . Sci. 259, 114063 (2025)

  7. [15]

    P.et al.Python materials genomics (pymatgen): A robust, open-source python library for materials analysis.Comput

    Ong, S. P.et al.Python materials genomics (pymatgen): A robust, open-source python library for materials analysis.Comput. Mater . Sci.68, 314–319 (2013). 28

  8. [16]

    & Grossman, J

    Xie, T. & Grossman, J. C. Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties.Phys. Rev. Lett.120, 145301 (2018)

  9. [17]

    & 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)

  10. [18]

    & DeCost, B

    Choudhary, K. & DeCost, B. Atomistic line graph neural network for improved materials property predictions.npj Comput. Mater .7, 185 (2021)

  11. [19]

    & 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)

  12. [20]

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

  13. [21]

    P., Simm, G

    Batatia, I., Kovács, D. P., Simm, G. N. C., Ortner, C. & Csányi, G. MACE: Higher order equivariant message passing neural networks for fast and accurate force fields. InAdvances in Neural Infor- mation Processing Systems, vol. 35 (2022). URL https://proceedings.neurips.cc/pape...

  14. [22]

    Merchant, A.et al.Scaling deep learning for materials discovery.Nature624, 80–85 (2023)

  15. [23]

    P.et al.Machine learning the voltage of electrode materials in metal-ion batteries.ACS Appl

    Joshi, R. P.et al.Machine learning the voltage of electrode materials in metal-ion batteries.ACS Appl. Mater . Interfaces11, 18494–18503 (2019)

  16. [24]

    Louis, S.-Y .et al.Accurate prediction of voltage of battery electrode materials using attention-based graph neural networks.ACS Appl. Mater . Interfaces14, 26587–26594 (2022)

  17. [25]

    & Masquelier, C

    Delacourt, C., Poizot, P., Morcrette, M., Tarascon, J.-M. & Masquelier, C. One-step low-temperature route for the preparation of electrochemically active LiMnPO4 powders.Chem. Mater .16, 93–99 (2004)

  18. [26]

    M., David, W

    Thackeray, M. M., David, W. I. F., Bruce, P. G. & Goodenough, J. B. Lithium insertion into manganese spinels.Mater . Res. Bull.18, 461–472 (1983)

  19. [27]

    & Hirai, T

    Ohzuku, T., Kitagawa, M. & Hirai, T. Electrochemistry of manganese dioxide in lithium nonaque- ous cell. III. X-ray diffractional study on the reduction of spinel-related manganese dioxide.J. Electrochem. Soc.137, 769–775 (1990). 29

  20. [28]

    A reflection on lithium-ion battery cathode chemistry.Nat

    Manthiram, A. A reflection on lithium-ion battery cathode chemistry.Nat. Commun.11, 1550 (2020)

  21. [29]

    Mater .35, 2210788 (2023)

    Schmidt, J.et al.Machine-learning-assisted determination of the global zero-temperature phase diagram of materials.Adv. Mater .35, 2210788 (2023)

  22. [30]

    Klimeš, J., Bowler, D. R. & Michaelides, A. Van der Waals density functionals applied to solids. Phys. Rev. B83, 195131 (2011)

  23. [31]

    & Furthmüller, J

    Kresse, G. & Furthmüller, J. Efficient iterative schemes forab initiototal-energy calculations using a plane-wave basis set.Phys. Rev. B54, 11169–11186 (1996)

  24. [32]

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

  25. [33]

    & Joubert, D

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

  26. [34]

    Yamada, A., Chung, S. C. & Hinokuma, K. Optimized LiFePO 4 for lithium battery cathodes.J. Electrochem. Soc.148, A224–A229 (2001)

  27. [35]

    K.et al.LiMnPO 4 as an advanced cathode material for rechargeable lithium batteries.J

    Martha, S. K.et al.LiMnPO 4 as an advanced cathode material for rechargeable lithium batteries.J. Electrochem. Soc.156, A541–A552 (2009)

  28. [36]

    Noh, H.-J., Youn, S., Yoon, C. S. & Sun, Y .-K. Comparison of the structural and electrochemical properties of layered Li[NixCoyMnz]O2 (x=1/3 , 0.5, 0.6, 0.7, 0.8 and 0.85) cathode material for lithium-ion batteries.J. Power Sources233, 121–130 (2013)

  29. [37]

    Reimers, J. N. & Dahn, J. R. Electrochemical andin situX-ray diffraction studies of lithium intercalation in LixCoO2.J. Electrochem. Soc.139, 2091–2097 (1992)

  30. [38]

    Atomgpt: Atomistic generative pretrained transformer for forward and inverse materials design.J

    Choudhary, K. Atomgpt: Atomistic generative pretrained transformer for forward and inverse materials design.J. Phys. Chem. Lett.15, 6909–6917 (2024)

  31. [39]

    Lee, J.et al.Agapi-agents: An open-access agentic ai platform for accelerated materials design on atomgpt.org.J. Phys. Chem. Lett.17, 7221–7231 (2026)

  32. [40]

    Slakonet: A unified slater–koster tight-binding framework using neural network infrastructure for the periodic table.J

    Choudhary, K. Slakonet: A unified slater–koster tight-binding framework using neural network infrastructure for the periodic table.J. Phys. Chem. Lett.16, 11109–11119 (2025). 30

  33. [41]

    P., Burke, K

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

  34. [42]

    L., Botton, G

    Dudarev, S. L., Botton, G. A., Savrasov, S. Y ., Humphreys, C. J. & Sutton, A. P. Electron-energy-loss spectra and the structural stability of nickel oxide: An LSDA+U study.Phys. Rev. B57, 1505–1509 (1998). 31

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.