Multitask learning with semiempirical orbital charges enables sample-efficient MLIPs

Ihor Neporozhnii; Oleksandr Voznyy; Sjoerd Hoogland

arxiv: 2605.24073 · v1 · pith:FUUIHUN3new · submitted 2026-05-22 · ⚛️ physics.chem-ph · cs.LG

Multitask learning with semiempirical orbital charges enables sample-efficient MLIPs

Ihor Neporozhnii , Sjoerd Hoogland , Oleksandr Voznyy This is my paper

Pith reviewed 2026-06-30 14:47 UTC · model grok-4.3

classification ⚛️ physics.chem-ph cs.LG

keywords machine learning interatomic potentialsmultitask learningsemiempirical orbital chargessample efficiencymaterials modelingequivariant neural networkselectronic structure

0 comments

The pith

Using semiempirical orbital charges in multitask training improves accuracy and sample efficiency of machine learning interatomic potentials.

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

The paper shows that adding the task of predicting orbitally resolved charges from a fast semiempirical method to the training of MLIPs leads to better energy predictions with less data. This multitask approach forces the model to learn a more physically grounded internal representation of the system. The charges come from calculations that are cheap and linear scaling, so they can be added to large datasets without much extra cost. The benefit is seen in lower errors and better generalization, and the model even organizes metals into chemically sensible groups on its own.

Core claim

Training MLIPs with both target energies and orbitally resolved GFN1-xTB charges via multitask learning yields a 46% lower energy mean absolute error and allows matching the performance of energy-only models with five times less training data. This method surpasses the use of more expensive DFT atomic charges and results in a latent space where metals cluster according to shared chemical properties. The orbital charges are used only in training, leaving inference speed unchanged.

What carries the argument

Multitask learning that jointly predicts energies and orbitally resolved semiempirical charges using an equivariant neural network architecture.

Load-bearing premise

The semiempirical orbital charges contain useful electronic information that is not corrupted by the approximations of the semiempirical method and that the model can use to improve its energy predictions.

What would settle it

If a model trained with the orbital charges shows higher energy errors than an energy-only model on a test set where the semiempirical charges differ markedly from accurate electronic structure calculations.

read the original abstract

Machine learning interatomic potentials (MLIPs) require generating computationally expensive, large-scale training datasets to accurately simulate materials and molecules. Incorporating electronic structure information using multitask learning improves sample efficiency, however, training on full Hamiltonian matrices, which scale quadratically with the number of atoms, is intractable for large datasets. In this work, we show that multitask learning utilizing orbitally resolved semiempirical charges significantly improves sample efficiency and accuracy in MLIPs. To efficiently predict orbital charges, we implement a specialized equivariant model, reducing charge prediction error compared to an invariant baseline. By augmenting training with computationally inexpensive GFN1-xTB orbital charges, which scale linearly with the number of atoms, our model achieves a 46\% reduction in energy mean absolute error and requires five times less data to match the performance of energy-only models. Furthermore, our approach outperforms models trained on expensive density functional theory (DFT) atomic charges, capturing orbitally resolved electronic complexity and forcing the network to learn a physically accurate latent space that spontaneously clusters metals by shared chemical properties. Because orbital charges are only required during training, this approach preserves inference efficiency, providing a scalable recipe for developing accurate, data-efficient foundation models for complex chemical systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Adding GFN1-xTB orbital charges as auxiliary targets gives a clear empirical lift to MLIP sample efficiency and beats DFT charges, but the mechanism and robustness need tighter checks.

read the letter

The main point is that multitask training on orbitally resolved GFN1-xTB charges cuts energy MAE by 46% and reaches the same performance with five times less data than energy-only models, while also outperforming models that use DFT atomic charges. They pair this with a specialized equivariant network for the charge prediction step.

What is actually new is the concrete choice of orbitally resolved semiempirical charges rather than atomic charges or full Hamiltonians, combined with the equivariant charge head. The linear scaling of GFN1-xTB makes the auxiliary data cheap to generate, and the charges are dropped at inference, so the efficiency gain is real on the training side.

The work does well on the practical angle: it directly targets the data-hungry nature of MLIPs for materials and molecules, and the claim that the model clusters metals by chemical properties is at least consistent with learning some electronic structure signal.

The soft spots are mostly around evidence strength. The abstract reports the headline numbers without error bars, dataset sizes, or split details, so it is hard to judge stability. The clustering observation is presented as support for physical transferability, but it could arise from the multitask loss or the equivariant architecture even if the xTB values carried no accurate chemistry; the text does not describe ablations with randomized targets or DFT-mismatched charges to rule that out. The outperformance over DFT charges is interesting if it holds, but it also raises the question of whether the semiempirical bias is helping or if something else is at play.

This paper is for groups working on data-efficient ML potentials who already use or can run xTB. It is worth sending to peer review because the empirical recipe is specific and the efficiency numbers, if reproducible, would matter for scaling foundation models.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces a multitask learning framework for machine learning interatomic potentials (MLIPs) that augments energy-only training with orbitally resolved charges computed via the GFN1-xTB semiempirical method. An equivariant architecture is used to predict these charges, and the authors report that incorporating the inexpensive auxiliary targets during training yields a 46% reduction in energy mean absolute error while requiring five times less data to reach the performance of baseline energy-only models. The approach is further claimed to outperform models trained on DFT atomic charges and to induce a latent space that spontaneously clusters metals according to chemical properties.

Significance. If the reported gains prove robust, the work supplies a practical route to data-efficient MLIPs by exploiting linear-scaling semiempirical electronic structure information only at training time. The distinction between orbitally resolved charges and conventional atomic charges, together with the explicit comparison to DFT charges, addresses a recognized limitation of prior multitask MLIP studies and could inform the design of foundation models for complex chemical systems.

major comments (2)

[Abstract] Abstract: the central performance claims (46% energy MAE reduction and 5× data reduction) are presented as aggregate numbers without error bars, training-set sizes, or explicit exclusion criteria for the datasets used in the comparison, rendering the statistical reliability and generalizability of the headline result impossible to assess from the given text.
[Abstract] Abstract: the claim that orbitally resolved GFN1-xTB charges force the network to learn a 'physically accurate latent space' is supported only by the observation of spontaneous metal clustering; no quantitative clustering metric, ablation against randomized or DFT-mismatched targets, or comparison of latent-space geometry is supplied to distinguish chemical transferability from multitask regularization effects.

minor comments (1)

[Abstract] Abstract: the statement that the equivariant charge model 'reduc[es] charge prediction error compared to an invariant baseline' is given without numerical values or reference to a table or figure that would allow the magnitude of the improvement to be evaluated.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments. We address each major comment point by point below.

read point-by-point responses

Referee: [Abstract] Abstract: the central performance claims (46% energy MAE reduction and 5× data reduction) are presented as aggregate numbers without error bars, training-set sizes, or explicit exclusion criteria for the datasets used in the comparison, rendering the statistical reliability and generalizability of the headline result impossible to assess from the given text.

Authors: We agree that the abstract, owing to length constraints, presents the headline metrics without accompanying statistical details. The main text (Sections 3.2 and 4.1) and supplementary information report the 46% MAE reduction and 5× data efficiency as averages over five random seeds with standard deviations, using explicitly sized subsets of the ANI-1x and custom periodic datasets (5k–100k structures) with documented outlier exclusion rules. We will revise the abstract to name the primary datasets and direct readers to the main text for the full statistics and error bars. revision: partial
Referee: [Abstract] Abstract: the claim that orbitally resolved GFN1-xTB charges force the network to learn a 'physically accurate latent space' is supported only by the observation of spontaneous metal clustering; no quantitative clustering metric, ablation against randomized or DFT-mismatched targets, or comparison of latent-space geometry is supplied to distinguish chemical transferability from multitask regularization effects.

Authors: The spontaneous clustering of metals by chemical families is offered as qualitative support for the claim. We acknowledge that this observation alone does not quantitatively separate the effect of orbital-charge multitask learning from generic regularization. In the revised manuscript we will add a silhouette-score analysis of the latent-space clusters together with an ablation that replaces the GFN1-xTB targets with randomized or DFT-mismatched values. revision: yes

Circularity Check

0 steps flagged

No significant circularity; results rely on external fixed targets

full rationale

The paper reports empirical gains (46% energy MAE reduction, 5× data efficiency) from multitask training on orbitally resolved charges produced by the fixed external GFN1-xTB code. These targets are independent of the MLIP parameters and are not derived from or fitted inside the present model. No equations, uniqueness theorems, or self-citations reduce the headline claims to definitions or fitted constants internal to the paper. The observed metal clustering is presented as a post-hoc observation rather than a load-bearing premise. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the premise that GFN1-xTB orbital charges encode useful electronic information that transfers to the MLIP without domain mismatch; no free parameters or invented entities are named in the abstract.

axioms (1)

domain assumption GFN1-xTB orbital charges capture transferable electronic structure information relevant to DFT-level energetics
Invoked to justify why the auxiliary task improves the latent space and outperforms DFT atomic charges.

pith-pipeline@v0.9.1-grok · 5758 in / 1292 out tokens · 39076 ms · 2026-06-30T14:47:30.738338+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

28 extracted references · 13 canonical work pages · 1 internal anchor

[1]

& G¨ unnemann, S.Gemnet: Universal directional graph neural networks for molecules(2021)

Gasteiger, J., Becker, F. & G¨ unnemann, S.Gemnet: Universal directional graph neural networks for molecules(2021)

2021
[2]

URL https://www

Musaelian, A.et al.Learning local equivariant representations for large-scale atomistic dynamics.Nature Communications14, 579 (2023). URL https://www. nature.com/articles/s41467-023-36329-y. Publisher: Nature Publishing Group

2023
[3]

P., Simm, G

Batatia, I., Kov´ acs, D. P., Simm, G. N. C., Ortner, C. & Cs´ anyi, G. MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields (2023). URL http://arxiv.org/abs/2206.07697. ArXiv:2206.07697 [stat]

work page arXiv 2023
[4]

M.et al.UMA: A Family of Universal Models for Atoms (2026)

Wood, B. M.et al.UMA: A Family of Universal Models for Atoms (2026). URL http://arxiv.org/abs/2506.23971. ArXiv:2506.23971 [cs]

work page arXiv 2026
[5]

S.et al.The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models (2026)

Levine, D. S.et al.The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models (2026). URL http://arxiv.org/abs/2505.08762. ArXiv:2505.08762 [physics]

work page arXiv 2025
[6]

K., Gao, M., Nam, J

Kaniselvan, M., Miller, B. K., Gao, M., Nam, J. & Levine, D. S.Learning from the electronic structure of molecules across the periodic table(2026). URL https: //openreview.net/forum?id=PS1YS8Wv4t

2026
[7]

URL https://www.nature.com/articles/s42256-023-00716-3

Deng, B.et al.CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling.Nature Machine Intelligence5, 1031–1041 (2023). URL https://www.nature.com/articles/s42256-023-00716-3. Publisher: Nature Publishing Group

2023
[8]

A., Hofstetter, A., Saha, S

Ghasemi, S. A., Hofstetter, A., Saha, S. & Goedecker, S. Interatomic potentials for ionic systems with density functional accuracy based on charge densities obtained by a neural network.Physical Review B92, 045131 (2015). URL https://link. aps.org/doi/10.1103/PhysRevB.92.045131. 10

work page doi:10.1103/physrevb.92.045131 2015
[9]

W., Finkler, J

Ko, T. W., Finkler, J. A., Goedecker, S. & Behler, J. A fourth-generation high- dimensional neural network potential with accurate electrostatics including non- local charge transfer.Nature Communications12, 398 (2021). URL https://www. nature.com/articles/s41467-020-20427-2. Publisher: Nature Publishing Group

2021
[10]

A., Sch¨ afer, M

Gubler, M., Finkler, J. A., Sch¨ afer, M. R., Behler, J. & Goedecker, S. Accelerating Fourth-Generation Machine Learning Potentials Using Quasi-Linear Scaling Par- ticle Mesh Charge Equilibration.Journal of Chemical Theory and Computation 20, 7264–7271 (2024). URL https://doi.org/10.1021/acs.jctc.4c00334. Publisher: American Chemical Society

work page doi:10.1021/acs.jctc.4c00334 2024
[11]

& Margraf, J

Vondr´ ak, M., Reuter, K. & Margraf, J. T. Pushing charge equilibration-based machine learning potentials to their limits.npj Computational Materials11, 288 (2025). URL https://www.nature.com/articles/s41524-025-01791-3. Publisher: Nature Publishing Group

2025
[12]

S., Leszczynski, J

Zubatyuk, R., Smith, J. S., Leszczynski, J. & Isayev, O. Accurate and trans- ferable multitask prediction of chemical properties with an atoms-in-molecules neural network.Science Advances5, eaav6490 (2019). URL https://www. science.org/doi/10.1126/sciadv.aav6490. Publisher: American Association for the Advancement of Science

work page doi:10.1126/sciadv.aav6490 2019
[13]

M., Zubatyuk, R

Anstine, D. M., Zubatyuk, R. & Isayev, O. AIMNet2: a neural network potential to meet your neutral, charged, organic, and elemental-organic needs.Chem- ical Science16, 10228–10244 (2025). URL https://pubs.rsc.org/en/content/ articlelanding/2025/sc/d4sc08572h. Publisher: The Royal Society of Chemistry

2025
[14]

& 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 and Computation13, 1989–2009 (2017). URL https://doi.org/10.1021/acs.jctc.7b001...

work page doi:10.1021/acs.jctc.7b00118 1989
[15]

Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds

Thomas, N.et al.Tensor field networks: Rotation- and translation-equivariant neural networks for 3d point clouds (2018). URL https://arxiv.org/abs/1802. 08219. arXiv:1802.08219

work page internal anchor Pith review Pith/arXiv arXiv 2018
[16]

URL https://doi.org/10.1038/s41467-022-29939-5

Batzner, S.et al.E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials.Nature Communications13, 2453 (2022). URL https://doi.org/10.1038/s41467-022-29939-5

work page doi:10.1038/s41467-022-29939-5 2022
[17]

& Zitnick, C

Passaro, S. & Zitnick, C. L. Krause, A.et al.(eds)Reducing SO(3) convolutions to SO(2) for efficient equivariant GNNs. (eds Krause, A.et al.)Proceedings of the 40th International Conference on Machine Learning, Vol. 202 ofProceedings of Machine Learning Research, 27420–27438 (2023). URL https://proceedings. mlr.press/v202/passaro23a.html. 11

2023
[18]

& Gomez-Bombarelli, R.Universally converging rep- resentations of matter across scientific foundation models(2025)

Edamadaka, S., Yang, S. & Gomez-Bombarelli, R.Universally converging rep- resentations of matter across scientific foundation models(2025). URL https: //openreview.net/forum?id=12ZCZVKm7r

2025
[19]

& Walsh, A.Platonic representation of foundation machine learning inter- atomic potentials(2026)

Li, Z. & Walsh, A.Platonic representation of foundation machine learning inter- atomic potentials(2026). URL https://openreview.net/forum?id=5Yt1eVV5gg

2026
[20]

& Großberger, L

McInnes, L., Healy, J., Saul, N. & Großberger, L. Umap: Uniform manifold approximation and projection.Journal of Open Source Software3, 861 (2018). URL https://doi.org/10.21105/joss.00861

work page doi:10.21105/joss.00861 2018
[21]

A recipe for charge density prediction(2024)

Fu, X.et al. A recipe for charge density prediction(2024). URL https: //openreview.net/forum?id=b7REKaNUTv

2024
[22]

& 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 Physics161, 062501 (2024). URL https://doi.org/10.1063/5.0216715

work page doi:10.1063/5.0216715 2024
[23]

& Smidt, T

Geiger, M. & Smidt, T. e3nn: Euclidean neural networks (2022). URL https: //arxiv.org/abs/2207.09453

work page arXiv 2022
[24]

Paszke, A.et al.Automatic differentiation in pytorch (2017)

2017
[25]

Fey, M.et al.PyG 2.0: Scalable Learning on Real World Graphs.Temporal Graph Learning Workshop @ KDD(2025)

2025
[26]

& The PyTorch Lightning team

Falcon, W. & The PyTorch Lightning team. PyTorch Lightning (2019). URL https://github.com/Lightning-AI/lightning

2019
[27]

Hydra - a framework for elegantly configuring complex applications

Yadan, O. Hydra - a framework for elegantly configuring complex applications. Github (2019). URL https://github.com/facebookresearch/hydra

2019
[28]

Goodfellow, A. S. & Nguyen, B. N. Graph-Based Internal Coordinate Analysis for Transition State Characterization.Journal of Chemical Theory and Computation 22, 2348–2357 (2026). URL https://doi.org/10.1021/acs.jctc.5c02073. Publisher: American Chemical Society. 12 Appendix A Equivariant prediction of orbital charges In Section 3, we introduced the orbital...

work page doi:10.1021/acs.jctc.5c02073 2026

[1] [1]

& G¨ unnemann, S.Gemnet: Universal directional graph neural networks for molecules(2021)

Gasteiger, J., Becker, F. & G¨ unnemann, S.Gemnet: Universal directional graph neural networks for molecules(2021)

2021

[2] [2]

URL https://www

Musaelian, A.et al.Learning local equivariant representations for large-scale atomistic dynamics.Nature Communications14, 579 (2023). URL https://www. nature.com/articles/s41467-023-36329-y. Publisher: Nature Publishing Group

2023

[3] [3]

P., Simm, G

Batatia, I., Kov´ acs, D. P., Simm, G. N. C., Ortner, C. & Cs´ anyi, G. MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields (2023). URL http://arxiv.org/abs/2206.07697. ArXiv:2206.07697 [stat]

work page arXiv 2023

[4] [4]

M.et al.UMA: A Family of Universal Models for Atoms (2026)

Wood, B. M.et al.UMA: A Family of Universal Models for Atoms (2026). URL http://arxiv.org/abs/2506.23971. ArXiv:2506.23971 [cs]

work page arXiv 2026

[5] [5]

S.et al.The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models (2026)

Levine, D. S.et al.The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models (2026). URL http://arxiv.org/abs/2505.08762. ArXiv:2505.08762 [physics]

work page arXiv 2025

[6] [6]

K., Gao, M., Nam, J

Kaniselvan, M., Miller, B. K., Gao, M., Nam, J. & Levine, D. S.Learning from the electronic structure of molecules across the periodic table(2026). URL https: //openreview.net/forum?id=PS1YS8Wv4t

2026

[7] [7]

URL https://www.nature.com/articles/s42256-023-00716-3

Deng, B.et al.CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling.Nature Machine Intelligence5, 1031–1041 (2023). URL https://www.nature.com/articles/s42256-023-00716-3. Publisher: Nature Publishing Group

2023

[8] [8]

A., Hofstetter, A., Saha, S

Ghasemi, S. A., Hofstetter, A., Saha, S. & Goedecker, S. Interatomic potentials for ionic systems with density functional accuracy based on charge densities obtained by a neural network.Physical Review B92, 045131 (2015). URL https://link. aps.org/doi/10.1103/PhysRevB.92.045131. 10

work page doi:10.1103/physrevb.92.045131 2015

[9] [9]

W., Finkler, J

Ko, T. W., Finkler, J. A., Goedecker, S. & Behler, J. A fourth-generation high- dimensional neural network potential with accurate electrostatics including non- local charge transfer.Nature Communications12, 398 (2021). URL https://www. nature.com/articles/s41467-020-20427-2. Publisher: Nature Publishing Group

2021

[10] [10]

A., Sch¨ afer, M

Gubler, M., Finkler, J. A., Sch¨ afer, M. R., Behler, J. & Goedecker, S. Accelerating Fourth-Generation Machine Learning Potentials Using Quasi-Linear Scaling Par- ticle Mesh Charge Equilibration.Journal of Chemical Theory and Computation 20, 7264–7271 (2024). URL https://doi.org/10.1021/acs.jctc.4c00334. Publisher: American Chemical Society

work page doi:10.1021/acs.jctc.4c00334 2024

[11] [11]

& Margraf, J

Vondr´ ak, M., Reuter, K. & Margraf, J. T. Pushing charge equilibration-based machine learning potentials to their limits.npj Computational Materials11, 288 (2025). URL https://www.nature.com/articles/s41524-025-01791-3. Publisher: Nature Publishing Group

2025

[12] [12]

S., Leszczynski, J

Zubatyuk, R., Smith, J. S., Leszczynski, J. & Isayev, O. Accurate and trans- ferable multitask prediction of chemical properties with an atoms-in-molecules neural network.Science Advances5, eaav6490 (2019). URL https://www. science.org/doi/10.1126/sciadv.aav6490. Publisher: American Association for the Advancement of Science

work page doi:10.1126/sciadv.aav6490 2019

[13] [13]

M., Zubatyuk, R

Anstine, D. M., Zubatyuk, R. & Isayev, O. AIMNet2: a neural network potential to meet your neutral, charged, organic, and elemental-organic needs.Chem- ical Science16, 10228–10244 (2025). URL https://pubs.rsc.org/en/content/ articlelanding/2025/sc/d4sc08572h. Publisher: The Royal Society of Chemistry

2025

[14] [14]

& 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 and Computation13, 1989–2009 (2017). URL https://doi.org/10.1021/acs.jctc.7b001...

work page doi:10.1021/acs.jctc.7b00118 1989

[15] [15]

Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds

Thomas, N.et al.Tensor field networks: Rotation- and translation-equivariant neural networks for 3d point clouds (2018). URL https://arxiv.org/abs/1802. 08219. arXiv:1802.08219

work page internal anchor Pith review Pith/arXiv arXiv 2018

[16] [16]

URL https://doi.org/10.1038/s41467-022-29939-5

Batzner, S.et al.E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials.Nature Communications13, 2453 (2022). URL https://doi.org/10.1038/s41467-022-29939-5

work page doi:10.1038/s41467-022-29939-5 2022

[17] [17]

& Zitnick, C

Passaro, S. & Zitnick, C. L. Krause, A.et al.(eds)Reducing SO(3) convolutions to SO(2) for efficient equivariant GNNs. (eds Krause, A.et al.)Proceedings of the 40th International Conference on Machine Learning, Vol. 202 ofProceedings of Machine Learning Research, 27420–27438 (2023). URL https://proceedings. mlr.press/v202/passaro23a.html. 11

2023

[18] [18]

& Gomez-Bombarelli, R.Universally converging rep- resentations of matter across scientific foundation models(2025)

Edamadaka, S., Yang, S. & Gomez-Bombarelli, R.Universally converging rep- resentations of matter across scientific foundation models(2025). URL https: //openreview.net/forum?id=12ZCZVKm7r

2025

[19] [19]

& Walsh, A.Platonic representation of foundation machine learning inter- atomic potentials(2026)

Li, Z. & Walsh, A.Platonic representation of foundation machine learning inter- atomic potentials(2026). URL https://openreview.net/forum?id=5Yt1eVV5gg

2026

[20] [20]

& Großberger, L

McInnes, L., Healy, J., Saul, N. & Großberger, L. Umap: Uniform manifold approximation and projection.Journal of Open Source Software3, 861 (2018). URL https://doi.org/10.21105/joss.00861

work page doi:10.21105/joss.00861 2018

[21] [21]

A recipe for charge density prediction(2024)

Fu, X.et al. A recipe for charge density prediction(2024). URL https: //openreview.net/forum?id=b7REKaNUTv

2024

[22] [22]

& 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 Physics161, 062501 (2024). URL https://doi.org/10.1063/5.0216715

work page doi:10.1063/5.0216715 2024

[23] [23]

& Smidt, T

Geiger, M. & Smidt, T. e3nn: Euclidean neural networks (2022). URL https: //arxiv.org/abs/2207.09453

work page arXiv 2022

[24] [24]

Paszke, A.et al.Automatic differentiation in pytorch (2017)

2017

[25] [25]

Fey, M.et al.PyG 2.0: Scalable Learning on Real World Graphs.Temporal Graph Learning Workshop @ KDD(2025)

2025

[26] [26]

& The PyTorch Lightning team

Falcon, W. & The PyTorch Lightning team. PyTorch Lightning (2019). URL https://github.com/Lightning-AI/lightning

2019

[27] [27]

Hydra - a framework for elegantly configuring complex applications

Yadan, O. Hydra - a framework for elegantly configuring complex applications. Github (2019). URL https://github.com/facebookresearch/hydra

2019

[28] [28]

Goodfellow, A. S. & Nguyen, B. N. Graph-Based Internal Coordinate Analysis for Transition State Characterization.Journal of Chemical Theory and Computation 22, 2348–2357 (2026). URL https://doi.org/10.1021/acs.jctc.5c02073. Publisher: American Chemical Society. 12 Appendix A Equivariant prediction of orbital charges In Section 3, we introduced the orbital...

work page doi:10.1021/acs.jctc.5c02073 2026