Pith. sign in

REVIEW 3 major objections 5 minor 93 references

OrbitAll: A Unified Quantum Mechanical Representation Deep Learning Framework for All Molecular Systems

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims one deep-learning architecture can represent any molecular system — charge, spin, or solvent — by learning corrections on top of cheap semi-empirical quantum orbitals, reaching chemical accuracy with about ten times less…

desk verdict A genuinely useful unified representation for charge/spin/solvent, with strong benchmarks, but the data-efficiency and solvent claims outrun the comparisons actually run. read the letter →

arxiv 2507.03853 v2 pith:IUVGJZJI submitted 2025-07-05 cs.LG physics.chem-ph

Beom Seok Kang , Vignesh C. Bhethanabotla , Amin Tavakoli , Maurice D. Hanisch , Arimitsu Horikawa-Strakovsky , Miguel Nouman , Danish Khan , William A. Goddard III
show 1 more author
Anima Anandkumar
This is my paper
classification cs.LGphysics.chem-ph
keywords quantumchemistrydeeplearningspin-polarizedorbitalfeaturesSE(3)-equivariantgraphneuralnetworkdelta-learningsemi-empiricaltight-bindingopen-shellmoleculesimplicitsolvation
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

OrbitAll is a bid to remove the biggest restriction on deep learning for chemistry: models that only work for neutral, closed-shell, gas-phase molecules. The paper argues that a single architecture can handle every molecular electronic state — radicals, ions, triplets, and molecules in solvents — if the input is not geometry alone but the converged electronic structure of a cheap quantum-mechanical calculation. Concretely, the framework runs the semi-empirical method spGFN1-xTB, extracts spin-polarized orbital features that automatically reflect charge, spin, and environment, and trains an SE(3)-equivariant graph neural network to predict only the energy gap between the cheap calculation and the target theory. On the paper's benchmarks this representation reaches chemical accuracy with roughly one tenth of the training data the best geometry-only model needs, predicts carbene spin gaps below 1 kcal/mol, and keeps near-chemical accuracy on solvated molecules and on polypeptides far larger than anything in the training set. A reader should care because this is a concrete path toward one foundation model for essentially all molecular quantum chemistry, at three to four orders of magnitude below DFT cost.

What carries the argument

The load-bearing object is the spin-polarized orbital feature set $T = (F_\alpha, F_\beta, P_\alpha, P_\beta, S, H_{\text{core}})$ — the $\alpha$- and $\beta$-spin Fock matrices, the $\alpha$- and $\beta$-spin density matrices, the overlap matrix, and the core Hamiltonian — generated by the semi-empirical tight-binding method spGFN1-xTB. These matrices, called QMMs, carry the physics by construction: integrals of the density matrices give the electron count and the spin excess, so charge and spin are encoded without any extra input channel, and the SCF solution responds to the environment (CPCM dielectric constants, external fields), so solvation is encoded as a perturbation of the converged mean field. Because each QMM block transforms under rotations by Wigner-D matrices, the whole representation is SE(3)-equivariant, which the network preserves: an E(3)-equivariant GNN built on the UNiTE architecture processes the matrices through diagonal reduction (reading intra-atomic block diagonals into atom embeddings) and block-wise convolution message passing (building inter-atomic messages from off-diagonal blocks with attention aggregation), then pools to the target property. The remaining ingredient is the delta-learning objective of predicting $\Delta y = y_{\text{target}} - y_{\text{spGFN1-xTB}}$, so the network only has to learn the correction from the cheap baseline to the expensive label.

What would settle it

The most direct check is to push the framework past what its inputs can represent: take a dication or dianion (formal charge ±2, which the paper says the charge-shift initialization does not support) with the same atoms as QM9star and evaluate the published model — a large error would confirm that 'all molecular systems' is bounded by the training charge distribution rather than by the orbital representation. A complementary test targets the load-bearing premise itself: on open-shell species where spGFN1-xTB is known to struggle (strongly correlated diradicals, late-transition-metal complexes), measure the SCF non-convergence rate and, for converged cases, compare the predicted spin gaps against the paper's QMSpin-level performance; if a substantial fraction of such molecules cannot generate features at all, or the network inherits the wrong spin ordering from the baseline, the unified-representation claim is falsified precisely where open-shell chemistry is hardest.

Watch

Extended reading notes

Core claim

The paper's central claim is that the information needed to distinguish any molecular system — its charge, its spin state, and its surroundings — is already present in the mean-field electronic structure of a cheap semi-empirical calculation, and that a neural network can be taught to read that information. OrbitAll uses spin-polarized (unrestricted) spGFN1-xTB to produce the feature set $T = (F_\alpha, F_\beta, P_\alpha, P_\beta, S, H_{\text{core}})$, the up- and down-spin Fock matrices, the corresponding density matrices, the overlap matrix, and the core Hamiltonian. These quantum mechanical matrices are SE(3)-equivariant by construction, the density matrices integrate to the electron count (charge) and to twice the total spin, and the converged Fock matrices shift with implicit solvation or applied electric fields, so one representation covers all conditions. The network then predicts the delta label $\Delta y = y_{\text{target}} - y_{\text{spGFN1-xTB}}$, learning only the correction from the semi-empirical baseline to the high-level target energy. Empirically the paper finds this scheme beats seven augmented geometry-based GNN baselines on charged and open-shell total energies and frontier orbital energies, achieves combined singlet/triplet carbene errors of 28.5 meV against a model that trains separate spin-state specialists, transfers across four implicit solvents with nearly identical errors, and extrapolates to 16–36 heavy-atom polypeptides at roughly $10^4$ speedup over DFT while staying near chemical accuracy.

Load-bearing premise

Everything depends on spGFN1-xTB converging and responding correctly to charge, spin, and solvent, because the orbital features are both the network's only input and the baseline it corrects; if the cheap quantum calculation fails or gives wrong physics (SCF non-convergence, wrong spin ordering, poor strong-correlation description), there is nothing sound left for the network to learn from.

Editorial extensions

If this is right

  • Joint training over species pays off: the single mixed-species model beats a spin-separated specialist on triplet carbene energies (29.5 vs 48.6 meV), implying one model can replace per-condition specialists.
  • A model that reaches chemical accuracy at roughly 7K training points makes very expensive labels affordable, so datasets computed at CCSD(T) or MRCISD levels become practical training targets.
  • Reliable extrapolation from 1–9 heavy atoms to 16–36 heavy atoms suggests the framework can be applied to systems where the reference DFT calculation is itself impractical.
  • One model trained on four implicit solvents gives nearly identical errors in each, supporting solvent-dependent reaction-path studies at roughly 100 times lower cost than explicit-solvation simulations with a competing universal model.
  • With a $10^3$–$10^4$ speedup over DFT and stability on charged species, the authors position OrbitAll for data-scarce tasks such as metal-organic framework band gaps and experimental property prediction.

Reading between the lines

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

  • The paper's 'all molecular systems' is best read as 'all systems the cheap SCF can describe': the charge-shift parameterization is acknowledged to fail for charges outside the training distribution (e.g., +2), so universality is a property of the representation plus training coverage, not of the architecture alone (editorial inference).
  • The sharpest measured gains concentrate where spin resolution matters — beta-spin frontier orbitals and triplet energies — which predicts that the representation's advantage will grow with spin polarization, such as on open-shell transition-metal complexes (editorial inference).
  • Because the semi-empirical backend cannot provide analytical gradients, OrbitAll is currently a property-prediction engine; coupling it to a differentiable tight-binding code would unlock forces and molecular dynamics, where the speedup becomes motion-scale (editorial inference).
  • The solvent results suggest a cheap route to environment-aware foundation models: since dielectric constants perturb the features directly, the same architecture should absorb many solvents without a separate environment-encoding subnetwork (editorial inference).
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 / 5 minor

Summary. OrbitAll introduces an SE(3)-equivariant graph neural network that consumes spin-polarized semi-empirical orbital features (Fα, Fβ, Pα, Pβ, S, Hcore) computed by spGFN1-xTB and predicts delta-labels relative to higher-level DFT targets. The framework is designed to handle molecules with different total charges, spin states, and implicit solvent environments, and is evaluated on QM9star (charged/open-shell species), QMSpin (singlet/triplet carbenes), Hessian QM9 (four implicit solvents), and a new polypeptide dataset for size extrapolation. The paper reports consistent accuracy improvements over augmented geometry-based baselines, claims a 10x data-efficiency advantage, and reports roughly 10^3--10^4 speedups over DFT.

Significance. If the central claims are fully supported, OrbitAll would be a valuable step toward a single machine-learning architecture applicable across charge states, spin multiplicities, and environments, with the practical benefit of data-efficient training and strong size extrapolation. The paper has clear strengths: an extensive benchmark suite with several consistently augmented baselines, ablations of architectural choices (Table S4), joint training on singlet/triplet carbenes that improves triplet and spin-gap prediction (Table 1), and an honest acknowledgment of limitations such as SCF convergence dependence and the lack of analytic gradients (Sections 3 and S4). The claims that need the most scrutiny are the 10x data-efficiency comparison, the solvent generalization evidence, and the title-level promise of 'all molecular systems.'

major comments (3)
  1. [Cost-Accuracy Analysis; Figures 3(A) and 4] The '10 times fewer training data' claim compares OrbitAll trained with delta-learning against direct-learning baselines only (e.g., DimeNet++-SC(D) in Figure 3(A)), while the paper also reports delta-learning baselines in Table S5 (e.g., EquiformerV2-SC(∆) reaches 9.86 meV all-species MAE at 400K versus OrbitAll's 8.50 meV). Because delta-learning alone is known to improve data efficiency, this comparison conflates the benefit of delta-learning with the benefit of the orbital-feature representation. The same issue affects the size-extrapolation comparison in Figure 4, which again uses direct-learning baselines. Please provide learning curves or interpolated data requirements for the -SC(∆) baselines, or explicitly reframe the claim as '10x fewer data than direct-learning geometry GNNs.' I also note that the charge-shift concern raised in review is mitigated because Section 4.2 states the charge shift b_Q was applied to all models, not only OrbitAll.
  2. [Section 2.3; Figure 5] The Hessian QM9 solvent experiment reports only OrbitAll's MAEs (8.8--9.5 meV across vacuum, toluene, THF, water) with no baseline comparison. There is no SchNet, PaiNN, or EquiformerV2 trained on the same four-solvent split, no direct-learning variant, and no ablation that simply concatenates the dielectric constant epsilon_r as an input feature to a state-of-the-art geometry GNN. As a result, the central claim of 'unified environment handling' is not actually benchmarked against any alternative; the experiment demonstrates that OrbitAll can be trained on solvent-labeled data and achieve low errors, but it does not establish that the orbital-feature mechanism outperforms simpler environment-conditioning approaches. Please add at least one strong baseline (ideally both direct- and delta-learning) and an epsilon_r-input baseline on the same train/validation/test split.
  3. [Abstract, Section 1, and Section S2.1] The abstract and introduction claim the framework handles 'arbitrary charges' and 'all molecular systems,' but Section S2.1 explicitly states that the charge-shift initialization strategy 'does not allow other charges than those in the training set distribution' and that extrapolation to out-of-distribution charges such as +2 or -2 is expected to fail. This is an internal inconsistency between the stated scope and an acknowledged limitation. Either temper the 'all'/'arbitrary' language in the title and abstract, or demonstrate the framework on a wider range of charge states (e.g., by training on data that includes dianions/dications). As written, the headline claim overstates the demonstrated capability.
minor comments (5)
  1. [Figures 2 and 3] The legend markers for direct-learning (D) versus delta-learning (∆) do not appear in the rendered figure text; please ensure the markers are visible and consistent in the final figures.
  2. [Equation (15)] The symbol p is used for parity in the diagonal reduction formula before it is explicitly defined; please define parity when it first appears.
  3. [Section 2.2 and Table S5] The text refers to DimeNet++-SC(D) as the 'next-best model' for the data-efficiency interpolation, but Table S5 shows EquiformerV2-SC(D) has a lower all-species MAE at 400K (15.43 meV versus 19.41 meV). Please clarify the criterion used to identify the next-best model, since the same table may otherwise appear inconsistent.
  4. [Section 2.2 vs Cost-Accuracy Analysis] Section 2.2 states that all models for total-energy prediction on QM9star were trained with delta-learning, while the Cost-Accuracy Analysis and Figure 3 use direct-learning baselines; please clarify in the main text which comparisons are delta-vs-delta and which are delta-vs-direct.
  5. [Introduction, 'first' claim] The claim that OrbitAll is 'the first deep learning-based method to simultaneously handle diverse charges, spin states, and environments' is not fully positioned against prior work: SpookyNet and TensorNet already handle charge and spin, and implicit-solvent or electric-field inputs have been used in other models. Please provide a more nuanced literature comparison to support the novelty claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: delta-learning, charge shifts, and the UNiTE backbone are disclosed, externally benchmarked, and applied consistently; the central predictions are tested on held-out DFT labels.

full rationale

The paper's central prediction is the total energy (or FMO levels) at a target level of theory, trained against external DFT labels and evaluated on held-out test sets, so there is no internal circularity in the core claim. The delta-learning strategy is explicit: Equation (3) defines the target as Δy = ytarget − ylow-level, and the final prediction adds the spGFN1-xTB energy back; this is a transparent physics-based baseline, not a fitted parameter disguised as a prediction. The learnable charge shift bₑ in Equation (21) is initialized from training-label averages, but it is applied to all models, including baselines, and it is a standard additive bias rather than a claim that per-charge offsets are derived from orbital features. The reliance on the authors' prior UNiTE backbone [35] is a self-citation, but UNiTE is a published, externally benchmarked architecture, and the paper does not invoke a uniqueness theorem or smuggle in an ansatz via that citation. The acknowledged limitation about SCF convergence (Section 3 and S4) does not create circularity; it is a practical dependence of the feature-generation pipeline. The solvent experiment lacks competing baselines, but absence of a comparison is a benchmarking limitation, not a circular derivation. No equation or design choice reduces the predicted quantity to the training labels by construction, so no specific circular step can be exhibited.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the fidelity of spGFN1-xTB features and the delta-learning assumption, both domain assumptions rather than first-principles guarantees. The free parameters are standard ML biases and hyperparameters, not fitted physical constants. No new physical entities are introduced.

free parameters (3)
  • Charge shift b_Q = Fitted per charge state (Q = -1, 0, +1) during training
    A learnable bias added in the final pooling (Eq. 21) to account for average energy differences between charge states; initialized from training set averages.
  • Element-wise energy biases b_Z = Fitted per element (H, C, N, O, etc.)
    Atom-type energy shifts initialized by linear regression of total energy to atomic numbers, then trained (Eq. 18).
  • Model hyperparameters (hidden dim 256, 4 layers, 8 channels, 8 heads) = As in Table S2
    Chosen by hand from UNiTE defaults; not derived from any principle beyond prior empirical success.
assumptions (4)
  • domain assumption spGFN1-xTB converges and provides physically meaningful spin-polarized orbital features for all target systems
    Section 2.1 states the low-level method 'can physically capture the major effects'; the framework fails when SCF does not converge, acknowledged in Discussion.
  • standard math QMMs transform as SE(3)-equivariant features under rotation and translation, and UNiTE preserves this equivariance
    Eq. (2) and Eq. (13) rely on Wigner-D matrix properties, which are standard group theory.
  • domain assumption Delta-learning between spGFN1-xTB and target theory yields a smoother, more learnable target for all species
    Section 2.1 and Figure S12 provide empirical support; it is a widely used strategy but not a theorem.
  • standard math Closed-shell singlets have Fα = Fβ and Pα = Pβ, unifying restricted and unrestricted representations
    Eqs. (7)-(8) in Section 4.1; standard UHF/RHF equivalence for closed-shell singlets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OrbitAll: A Unified Quantum Mechanical Representation Deep Learning Framework for All Molecular Systems." pith.science (2026). https://pith.science/paper/IUVGJZJI

@misc{pith2026250703853,
  author       = {Pith},
  title        = {Pith review of: OrbitAll: A Unified Quantum Mechanical Representation Deep Learning Framework for All Molecular Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IUVGJZJI}},
  note         = {Machine review of arXiv:2507.03853}
}
abstract

We introduce OrbitAll, a geometry- and physics-informed deep learning framework that encodes any molecular system with arbitrary charges, spins, and environmental effects using electronic structure information. It utilizes spin-polarized orbital features from the underlying quantum mechanical method and combines them with SE(3)-equivariant graph neural networks. OrbitAll demonstrates superior performance and generalization in predicting charged, open-shell, and solvated molecules, and robustly extrapolates to molecules significantly larger than the training data. OrbitAll achieves chemical accuracy using 10 times fewer training data than competing AI models, with approximately $10^3$ - $10^4$ speedup compared to density functional theory. Trained on a chemically diverse dataset, OrbitAll performs robustly on challenging molecular systems, and outperforms the foundational machine-learned interatomic potential, UMA, for highly charged species, despite using 35 times less molecular data and a 50-times-smaller model. After learning solvent effects, it accurately predicts solvent-dependent reaction pathways at about 100 times lower cost than explicit-solvation simulations using UMA.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

93 extracted references · 69 canonical work pages

  1. [1]

    J., Mori-S´ anchez, P

    Cohen, A. J., Mori-S´ anchez, P. & Yang, W. Challenges for density functional theory. Chemical reviews 112, 289–320 (2012). 18

    2012

  2. [2]

    Pereira, F. et al. Machine learning methods to predict density functional theory b3lyp energies of homo and lumo orbitals. Journal of chemical information and modeling 57, 11–21 (2017)

    2017

  3. [3]

    Mood, A. et al. Methyl anion affinities of the canonical organic functional groups. The Journal of organic chemistry 85, 4096–4102 (2020)

    2020

  4. [4]

    & Ohanessian, G

    Frison, G. & Ohanessian, G. A comparative study of semiempirical, ab initio, and dft methods in evaluating metal–ligand bond strength, proton affinity, and interactions between first and second shell ligands in zn-biomimetic complexes. Journal of computational chemistry 29, 416–433 (2008)

    2008

  5. [5]

    Density functional theory (Springer, 2011)

    Engel, E. Density functional theory (Springer, 2011)

    2011

  6. [6]

    Friesner, R. A. Ab initio quantum chemistry: Methodology and applications. Proceedings of the National Academy of Sciences 102, 6648–6653 (2005)

    2005

  7. [7]

    Fedik, N. et al. Extending machine learning beyond interatomic potentials for predicting molecular properties. Nature Reviews Chemistry 6, 653–672 (2022)

    2022

  8. [8]

    Fooshee, D. et al. Deep learning for chemical reaction prediction. Molecular Systems Design & Engineering 3, 442–452 (2018)

    2018

Show all 93 references
  1. [9]

    Musaelian, A. et al. Learning local equivariant representations for large-scale atomistic dynamics. Nature Communications 14, 579 (2023)

    2023

  2. [10]

    Qiao, Z., Welborn, M., Anandkumar, A., Manby, F. R. & Miller, T. F. OrbNet: Deep learning for quantum chemistry using symmetry-adapted atomic-orbital features. The Journal of Chemical Physics 153, 124111 (2020)

    2020

  3. [11]

    Christensen, A. S. et al. Orbnet denali: A machine learning potential for biological and organic chemistry with semi-empirical cost and dft accuracy. The Journal of Chemical Physics 155 (2021)

    2021

  4. [12]

    Batzner, S. et al. E (3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature communications 13, 2453 (2022)

    2022

  5. [13]

    Tavakoli, M., Chiu, Y. T. T., Carlton, A. M., Van Vranken, D. & Baldi, P. Chemi- cally informed deep learning for interpretable radical reaction prediction. Journal of Chemical Information and Modeling (2025)

    2025

  6. [14]

    & Baldi, P

    Tavakoli, M., Mood, A., Van Vranken, D. & Baldi, P. Quantum mechanics and machine learning synergies: graph attention neural networks to predict chemical reactivity. Journal of Chemical Information and Modeling 62, 2121–2132 (2022)

    2022

  7. [15]

    Unke, O. T. et al. SpookyNet: Learning force fields with electronic degrees of freedom and nonlocal effects. Nature Communications 12, 7273 (2021). URL https://www.nature.com/articles/s41467-021-27504-0. 19

    2021

  8. [16]

    P., Galvelis, R

    Simeon, G., Mirarchi, A., Pelaez, R. P., Galvelis, R. & De Fabritiis, G. Broaden- ing the scope of neural network potentials through direct inclusion of additional molecular attributes. Journal of Chemical Theory and Computation (2025)

    2025

  9. [17]

    T., Sauceda, H

    Sch¨ utt, K. T., Sauceda, H. E., Kindermans, P.-J., Tkatchenko, A. & M¨ uller, K.- R. Schnet–a deep learning architecture for molecules and materials. The Journal of Chemical Physics 148 (2018)

    2018

  10. [18]

    & Gastegger, M

    Sch¨ utt, K., Unke, O. & Gastegger, M. Meila, M. & Zhang, T. (eds) Equivariant message passing for the prediction of tensorial properties and molecular spectra . (eds Meila, M. & Zhang, T.) Proceedings of the 38th International Conference on Machine Learning, 9377–9388 (PMLR, 2021)

    2021

  11. [19]

    & G¨ unnemann, S

    Gasteiger, J., Groß, J. & G¨ unnemann, S. Directional message passing for molecular graphs. arXiv preprint arXiv:2003.03123 (2020)

    2020 arXiv

  12. [20]

    Faglioni, F., Merinov, B. V. & Goddard III, W. A. Room-temperature lithium phases from density functional theory. The Journal of Physical Chemistry C 120, 27104–27108 (2016)

    2016

  13. [21]

    Y., Merinov, B

    Das, T., Yang, M. Y., Merinov, B. V. & Goddard, W. A. Atomistic Simulations of Battery Materials and Processes , 13–76 (Springer International Publishing, Cham, 2024). URL https://doi.org/10.1007/978-3-031-47303-6 2

    2024 doi

  14. [22]

    T., Hoffmann, R., Stuyver, T

    Borden, W. T., Hoffmann, R., Stuyver, T. & Chen, B. Dioxygen: What makes this triplet diradical kinetically persistent? Journal of the American Chemical Soci- ety 139, 9010–9018 (2017). URL https://doi.org/10.1021/jacs.7b04232. PMID: 28613073

    2017 doi

  15. [23]

    Tavakoli, M., Chiu, Y. T. T., Baldi, P., Carlton, A. M. & Van Vranken, D. Rmechdb: A public database of elementary radical reaction steps. Journal of chemical information and modeling 63, 1114–1123 (2023)

    2023

  16. [24]

    Tavakoli, M. et al. Ai for interpretable chemistry: predicting radical mechanistic pathways via contrastive learning. Advances in Neural Information Processing Systems 36 (2024)

    2024

  17. [25]

    S., Bhethanabotla, V

    Kang, B. S., Bhethanabotla, V. C., Tavakoli, M., Goddard, W. & Anandkumar, A. Geometry aware deep learning for integrated closed-shell and open-shell systems (2024)

    2024

  18. [26]

    S., Tavakoli, M., Bhethanabotla, V

    Kang, B. S., Tavakoli, M., Bhethanabotla, V. C., Goddard III, W. A. & Anand- kumar, A. Orbnet-spin: Quantum mechanics informed geometric deep learning for open-shell systems (2024). URL https://ml4physicalsciences.github.io/2024/ files/NeurIPS ML4PS 2024 214.pdf. 20

    2024

  19. [27]

    R., Bursch, M., Grimme, S

    Maurer, L. R., Bursch, M., Grimme, S. & Hansen, A. Assessing density functional theory for chemically relevant open-shell transition metal reactions. Journal of Chemical Theory and Computation 17, 6134–6151 (2021)

    2021

  20. [28]

    & Kowalczyk, S

    Gutowski, M. & Kowalczyk, S. A study of free radical chemistry: their role and pathophysiological significance. Acta Biochimica Polonica 60, 1–16 (2013)

    2013

  21. [29]

    & Zhao, Y

    He, Q., Yu, B., Li, Z. & Zhao, Y. Density functional theory for battery materials. Energy & Environmental Materials 2, 264–279 (2019)

    2019

  22. [30]

    & Lim, C

    Dudev, T. & Lim, C. Competition among metal ions for protein binding sites: determinants of metal ion selectivity in proteins. Chemical reviews 114, 538–556 (2014)

    2014

  23. [31]

    Roberts, D. et al. Specific ion and buffer effects on protein–protein interactions of a monoclonal antibody. Molecular pharmaceutics 12, 179–193 (2015)

    2015

  24. [32]

    & Wu, C.-M

    Liang, X.-Y., Ding, N., Ng, S.-P. & Wu, C.-M. L. Adsorption of gas molecules on Ga-doped graphene and effect of applied electric field: A DFT study. Applied Surface Science 411, 11–17 (2017)

    2017

  25. [33]

    & Borgis, D

    Zhao, S., Ramirez, R., Vuilleumier, R. & Borgis, D. Molecular density functional theory of solvation: From polar solvents to water.The Journal of chemical physics 134 (2011)

    2011

  26. [34]

    & Salahub, D

    Wei, Proynov, E., Milet, A. & Salahub, D. Solvation of the hydroxide anion: A combined dft and molecular dynamics study. The Journal of Physical Chemistry A 104, 2384–2395 (2000)

    2000

  27. [35]

    Qiao, Z. et al. Informing geometric deep learning with electronic interactions to accelerate quantum chemistry. Proceedings of the National Academy of Sciences 119, e2205221119 (2022)

    2022

  28. [36]

    & von Lilienfeld, O

    Karandashev, K. & von Lilienfeld, O. A. An orbital-based representation for accurate quantum machine learning. The Journal of Chemical Physics 156 (2022)

    2022

  29. [37]

    Cheng, L., Welborn, M., Christensen, A. S. & Miller, T. F. A universal density matrix functional from molecular orbital-based machine learning: Transferability across organic molecules. The Journal of chemical physics 150 (2019)

    2019

  30. [38]

    Lu, F. et al. Fast near ab initio potential energy surfaces using machine learning. The Journal of Physical Chemistry A 126, 4013–4024 (2022)

    2022

  31. [39]

    B., Welborn, M

    Cheng, L., Kovachki, N. B., Welborn, M. & Miller III, T. F. Regression clustering for improved accuracy and training costs with molecular-orbital-based machine learning. Journal of Chemical Theory and Computation 15, 6668–6677 (2019). 21

    2019

  32. [40]

    & Miller Iii, T

    Cheng, L., Sun, J. & Miller Iii, T. F. Accurate molecular-orbital-based machine learning energies via unsupervised clustering of chemical space. Journal of Chemical Theory and Computation 18, 4826–4835 (2022)

    2022

  33. [41]

    E., Bhethanabotla, V

    Cheng, L., Sun, J., Deustua, J. E., Bhethanabotla, V. C. & Miller, T. F. Molecular-orbital-based machine learning for open-shell and multi-reference sys- tems with kernel addition Gaussian process regression. The Journal of Chemical Physics 157, 154105 (2022)

    2022

  34. [42]

    Venturella, C. et al. Unified deep learning framework for many-body quantum chemistry via green’s functions. Nature Computational Science (2025)

    2025

  35. [43]

    O., Rupp, M

    Ramakrishnan, R., Dral, P. O., Rupp, M. & von Lilienfeld, O. A. Big data meets quantum chemistry approximations: The ∆-machine learning approach. Journal of Chemical Theory and Computation 11, 2087–2096 (2015). URL https: //doi.org/10.1021/acs.jctc.5b00099. PMID: 26574412

    2015 doi

  36. [44]

    & Bjerrum, E

    Irwin, R., Dimitriadis, S., He, J. & Bjerrum, E. J. Chemformer: a pre- trained transformer for computational chemistry. Machine Learning: Science and Technology 3, 015022 (2022)

    2022

  37. [45]

    Fabrizio, A., Briling, K. R. & Corminboeuf, C. Spa hm: The spectrum of approximated hamiltonian matrices representations. Digital Discovery 1, 286–294 (2022)

    2022

  38. [46]

    R., Calvino Alonso, Y., Fabrizio, A

    Briling, K. R., Calvino Alonso, Y., Fabrizio, A. & Corminboeuf, C. Spahm (a, b): Encoding the density information from guess hamiltonian in quantum machine learning representations. Journal of chemical theory and computation 20, 1108– 1117 (2024)

    2024

  39. [47]

    & Grimme, S

    Neugebauer, H., B¨ adorf, B., Ehlert, S., Hansen, A. & Grimme, S. High- throughput screening of spin states for transition metal complexes with spin- polarized extended tight-binding methods. Journal of Computational Chemistry 44, 2120–2129 (2023). URL http://dx.doi.org/10.100...

    2023 doi

  40. [48]

    & Shushkov, P

    Grimme, S., Bannwarth, C. & Shushkov, P. A robust and accurate tight-binding quantum chemical method for structures, vibrational frequencies, and noncovalent interactions of large molecular systems parametrized for all spd-block elements (z = 1–86). Journal of Chemical Theory ...

    2017 doi

  41. [49]

    Liu, Y. et al. Spherical message passing for 3d graph networks (2022). URL https://arxiv.org/abs/2102.05013. arXiv:2102.05013

    2022 arXiv

  42. [50]

    & Smidt, T

    Liao, Y.-L., Wood, B., Das, A. & Smidt, T. EquiformerV2: Improved equivariant transformer for scaling to higher-degree representations (2024). URL https:// arxiv.org/abs/2306.12059. arXiv:2306.12059. 22

    2024 arXiv

  43. [51]

    & Hong, X

    Tang, M.-J., Zhu, T.-C., Zhang, S.-Q. & Hong, X. Qm9star, two million dft- computed equilibrium structures for ions and radicals with atomic information. Scientific Data 11, 1158 (2024)

    2024

  44. [52]

    Schwilk, M., Tahchieva, D. N. & von Lilienfeld, O. A. Large yet bounded: Spin gap ranges in carbenes. arXiv preprint arXiv:2004.10600 (2020)

    2020 arXiv

  45. [53]

    J., Kabalan, L., Stojanovic, L., Z´ olyomi, V

    Williams, N. J., Kabalan, L., Stojanovic, L., Z´ olyomi, V. & Pyzer-Knapp, E. O. Hessian qm9: A quantum chemistry database of molecular hessians in implicit solvents. Scientific data 12, 9 (2025)

    2025

  46. [54]

    Thomas, N. et al. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219 (2018)

    2018 arXiv

  47. [55]

    & Ostlund, N

    Szabo, A. & Ostlund, N. S. Modern Quantum Chemistry - Introduction to Advanced Electronic Structure Theory (Dover Publications, 1989)

    1989

  48. [56]

    & Schreiner, P

    Ruth, M., Gerbig, D. & Schreiner, P. R. Machine learning of coupled cluster (t)-energy corrections via delta ( δ)-learning. Journal of Chemical Theory and Computation 18, 4846–4855 (2022)

    2022

  49. [57]

    & Liu, F

    Chen, X., Li, P., Hruska, E. & Liu, F. δ-machine learning for quantum chemistry prediction of solution-phase molecular properties at the ground and excited states. Physical Chemistry Chemical Physics 25, 13417–13428 (2023)

    2023

  50. [58]

    G., Irle, S

    Zhu, J., Sumpter, B. G., Irle, S. et al. Artificial neural network correc- tion for density-functional tight-binding molecular dynamics simulations. MRS Communications 9, 867–873 (2019)

    2019

  51. [59]

    P., Simm, G., Ortner, C

    Batatia, I., Kovacs, D. P., Simm, G., Ortner, C. & Cs´ anyi, G. Mace: Higher order equivariant message passing neural networks for fast and accurate force fields. Advances in Neural Information Processing Systems 35, 11423–11436 (2022)

    2022

  52. [60]

    Gasteiger, J., Giri, S., Margraf, J. T. & G¨ unnemann, S. Fast and uncertainty- aware directional message passing for non-equilibrium molecules (2022). URL https://arxiv.org/abs/2011.14115. arXiv:2011.14115

    2022 arXiv

  53. [61]

    O., Rupp, M

    Ramakrishnan, R., Dral, P. O., Rupp, M. & Von Lilienfeld, O. A. Quantum chemistry structures and properties of 134 kilo molecules. Scientific Data 1, 140022 (2014)

    2014

  54. [62]

    Gastegger, M., Sch¨ utt, K. T. & M¨ uller, K.-R. Machine learning of solvent effects on molecular spectra and reactions. Chemical science 12, 11473–11483 (2021)

    2021

  55. [63]

    Ward, L. et al. Graph-based approaches for predicting solvation energy in multiple solvents: open datasets and machine learning models. The Journal of Physical Chemistry A 125, 5990–5998 (2021). 23

    2021

  56. [64]

    Falletta, S. et al. Unified differentiable learning of electric response. Nature Communications 16, 4031 (2025)

    2025

  57. [65]

    & Cossi, M

    Barone, V. & Cossi, M. Quantum calculation of molecular energies and energy gradients in solution by a conductor solvent model. The Journal of Physical Chemistry A 102, 1995–2001 (1998)

    1998

  58. [66]

    & Houk, K

    Takano, Y. & Houk, K. Benchmarking the conductor-like polarizable continuum model (cpcm) for aqueous solvation free energies of neutral and ionic organic molecules. Journal of Chemical Theory and Computation 1, 70–77 (2005)

    2005

  59. [67]

    & Hartke, B

    Alibakhshi, A. & Hartke, B. Implicitly perturbed hamiltonian as a class of versa- tile and general-purpose molecular representations for machine learning. Nature communications 13, 1245 (2022)

    2022

  60. [68]

    Rosen, A. S. et al. Machine learning the quantum-chemical properties of metal– organic frameworks for accelerated materials discovery. Matter 4, 1578–1597 (2021)

    2021

  61. [69]

    Rosen, A. S. et al. High-throughput predictions of metal–organic framework electronic properties: theoretical challenges, graph neural networks, and data exploration. npj Computational Materials 8, 1–10 (2022)

    2022

  62. [70]

    Ganscha, S. et al. The qcml dataset, quantum chemistry reference data from 33.5 m dft and 14.7 b semi-empirical calculations. Scientific Data 12, 406 (2025)

    2025

  63. [71]

    Levine, D. S. et al. The open molecules 2025 (omol25) dataset, evaluations, and models (2025). URL https://arxiv.org/abs/2505.08762. arXiv:2505.08762

    2025

  64. [72]

    Wood, B. M. et al. Uma: A family of universal models for atoms (2025). URL https://arxiv.org/abs/2506.23971. arXiv:2506.23971

    2025

  65. [73]

    & Grimme, S

    Pracht, P., Caldeweyher, E., Ehlert, S. & Grimme, S. A robust non-self-consistent tight-binding quantum chemistry method for large molecules. ChemRxiv (2019)

    2019

  66. [74]

    awvwgk/tblite: Light-weight tight-binding framework

    Ehlert, S. awvwgk/tblite: Light-weight tight-binding framework. https://github. com/awvwgk/tblite (2024)

    2024

  67. [75]

    & Grimme, S

    Bannwarth, C., Ehlert, S. & Grimme, S. Gfn2-xtb—an accurate and broadly parametrized self-consistent tight-binding quantum chemical method with mul- tipole electrostatics and density-dependent dispersion contributions. Journal of chemical theory and computation 15, 1652–1671 (2019)

    2019

  68. [76]

    & Grimme, S

    Froitzheim, T., M¨ uller, M., Hansen, A. & Grimme, S. g-xtb: A general-purpose extended tight-binding electronic structure method for the elements h to lr (z= 1–103). ChemRxiv (2025). 24

    2025

  69. [77]

    K., Otero-de La-Roza, A

    Prasad, V. K., Otero-de La-Roza, A. & DiLabio, G. A. Pepconf, a diverse data set of peptide conformational energies. Scientific data 6, 1–9 (2019)

    2019

  70. [78]

    V., Cramer, C

    Marenich, A. V., Cramer, C. J. & Truhlar, D. G. Universal solvation model based on solute electron density and on a continuum model of the solvent defined by the bulk dielectric constant and atomic surface tensions. The Journal of Physical Chemistry B 113, 6378–6396 (2009)

    2009

  71. [79]

    Kingma, D. P. & Ba, J. Adam: A method for stochastic optimization (2015). URL https://arxiv.org/abs/1412.6980

    2015 arXiv

  72. [80]

    J., Devlin, F

    Stephens, P. J., Devlin, F. J., Chabalowski, C. F. & Frisch, M. J. Ab initio calculation of vibrational absorption and circular dichroism spectra using density functional force fields. The Journal of physical chemistry 98, 11623–11627 (1994)

    1994

  73. [81]

    Becke, A. D. Density-functional exchange-energy approximation with correct asymptotic behavior. Physical review A 38, 3098 (1988)

    1988

  74. [82]

    & Parr, R

    Lee, C., Yang, W. & Parr, R. G. Development of the colle-salvetti correlation- energy formula into a functional of the electron density. Physical review B 37, 785 (1988)

    1988

  75. [83]

    Liu, S. et al. Symmetry-informed geometric representation for molecules, proteins, and crystalline materials. Advances in neural information processing systems 36, 66084–66101 (2023)

    2023

  76. [84]

    & Piquemal, J.-P

    Pl´ e, T., Lagard` ere, L. & Piquemal, J.-P. Force-field-enhanced neural network interactions: from local equivariant embedding to atom-in-molecule properties and long-range effects. Chemical Science 14, 12554–12569 (2023)

    2023

  77. [85]

    Caldeweyher, E. et al. A generally applicable atomic-charge dependent london dispersion correction. The Journal of chemical physics 150 (2019)

    2019

  78. [86]

    Software update: The orca program system—version 6.0

    Neese, F. Software update: The orca program system—version 6.0. Wiley Interdisciplinary Reviews: Computational Molecular Science 15, e70019 (2025)

    2025

  79. [87]

    & Krieg, H

    Grimme, S., Antony, J., Ehrlich, S. & Krieg, H. A consistent and accurate ab initio parametrization of density functional dispersion correction (dft-d) for the 94 elements h-pu. The Journal of chemical physics 132 (2010)

    2010

  80. [88]

    Smith, D. G. et al. Psi4 1.4: Open-source software for high-throughput quantum chemistry. The Journal of chemical physics 152 (2020)

    2020

  81. [89]

    St. John, P. C. et al. Quantum chemical calculations for over 200,000 organic radical species and 40,000 associated closed-shell molecules. Scientific Data 7, 244 (2020). 25

    2020

  82. [90]

    & Riniker, S

    B¨ oselt, L., Th¨ urlemann, M. & Riniker, S. Machine learning in qm/mm molecular dynamics simulations of condensed-phase systems. Journal of Chemical Theory and Computation 17, 2641–2658 (2021)

    2021

  83. [91]

    & Ceriotti, M

    Bigi, F., Langer, M. & Ceriotti, M. The dark side of the forces: assessing non- conservative force models for atomistic machine learning (2025). URL https: //arxiv.org/abs/2412.11569. arXiv:2412.11569

    2025

  84. [92]

    Qiao, Z. et al. Multi-task learning for electronic structure to predict and explore molecular potential energy surfaces. arXiv preprint arXiv:2011.02680 (2020)

    2020 arXiv

  85. [93]

    & Grimme, S

    Friede, M., H¨ olzer, C., Ehlert, S. & Grimme, S. dxtb—an efficient and fully differentiable framework for extended tight-binding. The Journal of Chemical Physics 161 (2024). 26 S1 Architecture Creating Messages by Block-wise Convolutions The block-wise convolutions characteri...

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.