pith. sign in

arxiv: 2606.09647 · v1 · pith:J4XNFIFVnew · submitted 2026-06-08 · ❄️ cond-mat.mtrl-sci · physics.comp-ph

Ab initio parametrization of distributed polarizable force fields

Felix Post , Jean-Philip Filling , Toulik Maitra , Falk May , Michael Wand , Denis Andrienko This is my paper

Pith reviewed 2026-06-27 15:41 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci physics.comp-ph
keywords polarizable force fieldsatomic polarizabilitygraph neural networkab initio parametrizationmolecular polarizability tensorsdistributed charge modelselectronic response properties
0
0 comments X

The pith

Assigning polarizability to individual atoms as tensors rather than atom-type scalars improves force-field accuracy for ions and excited states.

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

The paper shows that polarizable force fields become more accurate and broadly applicable when each atom receives its own polarizability value instead of sharing one with all atoms of the same type, and when that value is a full tensor rather than a single number. This change lets the models describe cations, anions, and electronically excited molecules that classical approaches handle poorly. The authors derive the tensors directly from quantum calculations on small molecules and train a graph neural network to predict them for new structures, removing the need for expensive calculations during large-scale simulations while adding a built-in check for when the model breaks down.

Core claim

Atomic polarizability tensors assigned to individual atoms, obtained from ab initio calculations on small organic molecules and predicted by a message-passing graph neural network, produce distributed polarizable force fields that describe neutral molecules more accurately and extend reliably to cations, anions, and excited states, all without adding cost to the molecular-dynamics step.

What carries the argument

Message-passing graph neural network that learns to map molecular graphs to per-atom polarizability tensors and scalars derived from first-principles calculations.

If this is right

  • Force fields can now be used for simulations involving charged species and photo-excited molecules without separate parametrization schemes.
  • Electronic response properties such as refractive index become accessible in large-scale molecular-dynamics runs at classical cost.
  • A diagnostic based on the difference between predicted and reference polarizabilities flags molecules where the model is likely to fail.
  • The same workflow applies to any conjugated organic building block once the network is trained on a representative set of small molecules.

Where Pith is reading between the lines

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

  • The per-atom tensor approach may reduce the need for manual atom-type assignment when building models for new materials libraries.
  • Because the network adds no runtime cost, the method could be inserted into existing polarizable simulation packages with minimal code changes.
  • Extending the training set to include inorganic ions or metal-organic complexes would test whether the same transferability holds beyond organic molecules.

Load-bearing premise

Polarizability values calculated on small molecules can be transferred by the neural network to larger or charged molecules without introducing large systematic errors.

What would settle it

Direct comparison of the neural-network-predicted polarizabilities against new ab initio calculations for a molecule outside the training set that contains an ion or an excited-state geometry, showing whether the error remains below the threshold needed for accurate refractive-index or density-of-states predictions.

Figures

Figures reproduced from arXiv: 2606.09647 by Denis Andrienko, Falk May, Felix Post, Jean-Philip Filling, Michael Wand, Toulik Maitra.

Figure 1
Figure 1. Figure 1: A representative subset of typical molecules from the molecular dataset is shown. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Representative subset of neutral molecules orderd and selected according their sizes. The diatom CO being [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Benchmark of Thole-model molecular polarizabilities for 76 [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: [λ = 0] Benchmark of Thole-model molecular polarizabilities for 75 singly charged cations (mCBP excluded) against DFT reference data. Layout and symbols as in [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: [λ = 0] Benchmark of Thole-model molecular polarizabilities for 76 singly charged anions against DFT reference data. Layout and symbols as in [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of predicted versus DFT-calculated molecular polarizabilities for the first singlet excited state [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of predicted versus DFT-calculated molecular polarizabilities for the first triplet excited state [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Conformational transferability of fitted atomic polarizabilities for TPBi. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison of experimental and simulated [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Atom-wise test MAE of the GNN-predicted atom-resolved polarizabilities for neutral, charged, and excited-state datasets. Each point corresponds to one independent train/validation/test split; black markers indicate the mean and standard deviation over all runs. puted by a learned edge network ϕe, m (l) ij = ϕe  h (l) i , h (l) j , eij ⊙ g (l) ij . (6) Here, h (l) i and h (l) j are the scalar node featur… view at source ↗
read the original abstract

Polarizable force fields offer superior transferability and accuracy compared to classical force fields, enabling access to electronic response properties such as refractive index and electronic density of states. Here, we demonstrate two key improvements that significantly enhance their accuracy: (1) assigning atomic polarizability to individual atoms rather than atom types, and (2) employing atomic polarizability tensors instead of scalar values. These modifications extend the applicability of polarizable force fields to cations, anions, and excited states, while also providing more accurate descriptions of neutral molecules. We propose a first-principles-based parameterization procedure for atomic polarizability tensors and scalars, validated on a set of small organic molecules with conjugated building blocks. To overcome the computational cost of ab initio calculations, we train a message-passing graph neural network to predict polarizability parameters, enabling efficient and scalable parameterization. Crucially, this approach imposes no additional computational cost during simulations and provides a clear diagnostic criterion for identifying cases where polarizable force field models fail to accurately describe molecular polarizability.

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.

Referee Report

2 major / 2 minor

Summary. The paper claims that assigning per-atom polarizability tensors (rather than atom-type scalars) derived from ab initio calculations on small neutral organic molecules, then predicted via a message-passing GNN, yields more accurate distributed polarizable force fields with no extra simulation cost; this is asserted to extend applicability to cations, anions, and excited states while improving neutral-molecule descriptions, with a diagnostic for model failure.

Significance. If the transferability claims hold, the work would advance first-principles parametrization of polarizable force fields by replacing atom-type approximations with atom-specific tensors and enabling scalable GNN-based assignment; the absence of added computational cost during MD and the diagnostic criterion are concrete strengths.

major comments (2)
  1. [Abstract and §4] Abstract and §4 (validation): the central claim that the per-atom tensor approach 'extends the applicability ... to cations, anions, and excited states' is unsupported because the reported ab initio reference data and GNN predictions are confined to neutral ground-state molecules with conjugated blocks; no separate calculations or transfer tests for charged or electronically excited systems are provided, leaving the extension as an assumption rather than a demonstrated result.
  2. [§3] §3 (GNN training): while the GNN is trained on ab initio polarizabilities from small molecules, the manuscript does not report error metrics or transferability tests on systems outside the neutral training distribution, which is required to substantiate the broader applicability asserted in the abstract.
minor comments (2)
  1. [§2] Notation for the polarizability tensor components should be clarified with an explicit equation relating the distributed atomic tensors to the molecular polarizability.
  2. [Figure 2] Figure captions for the GNN architecture and validation plots should include the exact training/validation split sizes and the molecules used.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We agree that the claims regarding extension to cations, anions, and excited states are not supported by explicit calculations in the current work and must be revised to reflect the demonstrated scope on neutral ground-state molecules. We will update the abstract, §3, and §4 accordingly while preserving the core technical contributions on per-atom tensors and the GNN predictor.

read point-by-point responses

  1. Referee: [Abstract and §4] Abstract and §4 (validation): the central claim that the per-atom tensor approach 'extends the applicability ... to cations, anions, and excited states' is unsupported because the reported ab initio reference data and GNN predictions are confined to neutral ground-state molecules with conjugated blocks; no separate calculations or transfer tests for charged or electronically excited systems are provided, leaving the extension as an assumption rather than a demonstrated result.

    Authors: We acknowledge that the manuscript provides no ab initio data or GNN predictions for charged or excited systems, so the extension claim is prospective rather than demonstrated. The per-atom tensor formulation is motivated by the expectation that local electronic response (captured via ab initio tensors) will adapt to changes in charge or electronic state where fixed atom-type scalars cannot, but this remains an untested hypothesis here. We will revise the abstract and §4 to state that the approach is designed to enable such extensions and that explicit validation on cations, anions, and excited states is planned for future work. revision: yes

  2. Referee: [§3] §3 (GNN training): while the GNN is trained on ab initio polarizabilities from small molecules, the manuscript does not report error metrics or transferability tests on systems outside the neutral training distribution, which is required to substantiate the broader applicability asserted in the abstract.

    Authors: The GNN is trained and evaluated exclusively on the neutral small-molecule dataset; no error metrics or tests on out-of-distribution systems (charged species, excited states, or larger molecules) are reported. We agree this limits substantiation of the broader applicability stated in the abstract. We will revise §3 to explicitly delimit the current training distribution, add a discussion of expected limitations for out-of-distribution cases, and reference the diagnostic criterion already in the manuscript as a practical safeguard for new systems. The abstract will be updated to match. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation uses external ab initio references

full rationale

The parameterization procedure starts from ab initio calculations on small neutral organic molecules to obtain per-atom polarizability tensors and scalars, then trains a message-passing GNN on those derived values to enable prediction for new molecules. No step reduces a claimed prediction or first-principles result to its own inputs by construction, nor does any load-bearing premise rest on a self-citation chain. The extension to cations, anions, and excited states is asserted on the basis of the per-atom tensor representation but is not validated within the described neutral-molecule test set; this constitutes an untested transferability assumption rather than circularity. The overall chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on ab initio calculations serving as reliable ground truth for polarizabilities and on the GNN successfully generalizing those values; no invented entities are introduced.

free parameters (1)
  • GNN model parameters and hyperparameters
    Trained on ab initio-derived polarizability data to enable prediction for new molecules.
axioms (1)
  • domain assumption Ab initio quantum calculations on small organic molecules provide transferable reference values for atomic polarizability tensors and scalars
    Invoked as the basis for the first-principles parameterization procedure.

pith-pipeline@v0.9.1-grok · 5716 in / 1179 out tokens · 27407 ms · 2026-06-27T15:41:21.527163+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 22 canonical work pages

  1. [2]

    Physical Chemistry Chemical Physics , author =

    Using atomic charges to model molecular polarization , volume =. Physical Chemistry Chemical Physics , author =. 2022 , note =. doi:10.1039/D1CP03542H , abstract =

    work page doi:10.1039/d1cp03542h 2022

  2. [3]

    The Journal of Physical Chemistry A , author =

    Ambiguities in. The Journal of Physical Chemistry A , author =. 2024 , note =. doi:10.1021/acs.jpca.4c01890 , abstract =

    work page doi:10.1021/acs.jpca.4c01890 2024

  3. [4]

    Journal of Chemical Theory and Computation , author =

    Unifying. Journal of Chemical Theory and Computation , author =. 2023 , note =. doi:10.1021/acs.jctc.3c00341 , abstract =

    work page doi:10.1021/acs.jctc.3c00341 2023

  4. [5]

    The journal of physical chemistry

    A. The journal of physical chemistry. A , author =. 2009 , pmid =. doi:10.1021/jp808952m , abstract =

    work page doi:10.1021/jp808952m 2009

  5. [6]

    2007 , keywords =

    Chemical Physics Letters , author =. 2007 , keywords =. doi:10.1016/j.cplett.2007.02.065 , abstract =

    work page doi:10.1016/j.cplett.2007.02.065 2007

  6. [8]

    Annual Review of Biophysics , author =

    Polarizable. Annual Review of Biophysics , author =. 2019 , keywords =. doi:10.1146/annurev-biophys-070317-033349 , abstract =

    work page doi:10.1146/annurev-biophys-070317-033349 2019

  7. [9]

    Advanced Energy Materials , volume =

    Scherer, Christoph and Kinaret, Naomi and Lin, Kun-Han and Qaisrani, Muhammad Nawaz and Post, Felix and May, Falk and Andrienko, Denis , title =. Advanced Energy Materials , volume =. doi:https://doi.org/10.1002/aenm.202403124 , url =. https://advanced.onlinelibrary.wiley.com/doi/pdf/10.1002/aenm.202403124 , year =

    work page doi:10.1002/aenm.202403124

  8. [10]

    M. J. Frisch and G. W. Trucks and H. B. Schlegel and G. E. Scuseria and M. A. Robb and J. R. Cheeseman and G. Scalmani and V. Barone and G. A. Petersson and H. Nakatsuji and X. Li and M. Caricato and A. V. Marenich and J. Bloino and B. G. Janesko and R. Gomperts and B. Mennucci and H. P. Hratchian and J. V. Ortiz and A. F. Izmaylov and J. L. Sonnenberg an...

    2016

  9. [11]

    2026 , eprint=

    Vibrational infrared and Raman spectra of the methanol molecule with equivariant neural-network property surfaces , author=. 2026 , eprint=

    2026

  10. [12]

    Long , title =

    Derek A. Long , title =. 2002 , publisher =. doi:10.1002/0470845767 , URL =

    work page doi:10.1002/0470845767 2002

  11. [13]

    Zeitschrift für Physik , volume =

    George Placzek , title =. Zeitschrift für Physik , volume =. 1931 , doi =

    1931

  12. [14]

    Peach, Michael J. G. and Williamson, Matthew J. and Tozer, David J. , title =. Journal of Chemical Theory and Computation , volume =. 2011 , doi =

    2011

  13. [15]

    The Journal of Chemical Physics , volume =

    Dreuw, Andreas and Weisman, Jennifer L. and Head-Gordon, Martin , title =. The Journal of Chemical Physics , volume =. 2003 , month =. doi:10.1063/1.1590951 , url =

    work page doi:10.1063/1.1590951 2003

  14. [16]

    Physical Chemistry Chemical Physics , author =

    Evaluating excited state atomic polarizabilities of chromophores , volume =. Physical Chemistry Chemical Physics , author =. 2018 , keywords =. doi:10.1039/C7CP08549D , number =

    work page doi:10.1039/c7cp08549d 2018

  15. [17]

    The Journal of Chemical Physics , author =

    Polarizability of molecular clusters as calculated by a dipole interaction model , volume =. The Journal of Chemical Physics , author =. 2002 , pages =. doi:10.1063/1.1433747 , language =

    work page doi:10.1063/1.1433747 2002

  16. [18]

    Journal of Chemical Theory and Computation , volume =

    Liang, Jiashu and Feng, Xintian and Hait, Diptarka and Head-Gordon, Martin , title =. Journal of Chemical Theory and Computation , volume =. 2022 , doi =

    2022

  17. [19]

    and Limas, Nidia Gabaldon , title =

    Manz, Thomas A. and Limas, Nidia Gabaldon , title =. RSC Adv. , year =. doi:10.1039/C6RA04656H , url =

    work page doi:10.1039/c6ra04656h

  18. [20]

    , title =

    Limas, Nidia Gabaldon and Manz, Thomas A. , title =. RSC Adv. , year =. doi:10.1039/C6RA05507A , url =

    work page doi:10.1039/c6ra05507a

  19. [21]

    , title =

    Limas, Nidia Gabaldon and Manz, Thomas A. , title =. RSC Adv. , year =. doi:10.1039/C7RA11829E , url =

    work page doi:10.1039/c7ra11829e

  20. [22]

    Zhang, T

    Martin H. Müser and Sergey V. Sukhomlinov and Lars Pastewka , title =. Advances in Physics: X , volume =. 2023 , publisher =. doi:10.1080/23746149.2022.2093129 , url =

    work page doi:10.1080/23746149.2022.2093129 2023

  21. [23]

    Journal of Chemical Theory and Computation , author =

    Polarizable. Journal of Chemical Theory and Computation , author =. 2011 , pages =. doi:10.1021/ct200304d , language =

    work page doi:10.1021/ct200304d 2011

  22. [24]

    Versatile Object-Oriented Toolkit for Coarse-Graining Applications , journal =

    R. Versatile Object-Oriented Toolkit for Coarse-Graining Applications , journal =. 2009 , doi =

    2009

  23. [25]

    1981 , issn =

    Molecular polarizabilities calculated with a modified dipole interaction , journal =. 1981 , issn =. doi:https://doi.org/10.1016/0301-0104(81)85176-2 , url =

    work page doi:10.1016/0301-0104(81)85176-2 1981

  24. [26]

    1971 , journal =

    Iterative solution of large linear systems , author =. 1971 , journal =

    1971

  25. [27]

    Atom dipole interaction model for molecular polarizability

    Applequist, Jon and Carl, James R and Fung, Kwok-Kueng , year =. Atom dipole interaction model for molecular polarizability. Journal of the American Chemical Society , number =

  26. [28]

    The Journal of Chemical Physics , author =

    Excited state polarizabilities of conjugated molecules calculated using time dependent density functional theory , volume =. The Journal of Chemical Physics , author =. 2001 , pages =. doi:10.1063/1.1415085 , abstract =

    work page doi:10.1063/1.1415085 2001

  27. [29]

    and Liu, Shubin , month = jan, year =

    Zhao, Dongbo and He, Xin and Ayers, Paul W. and Liu, Shubin , month = jan, year =. Excited-. Molecules , publisher =. doi:10.3390/molecules28062576 , abstract =

    work page doi:10.3390/molecules28062576

  28. [30]

    , month = dec, year =

    Ayers, Paul W. , month = dec, year =. The physical basis of the hard/soft acid/base principle , volume =. Faraday Discussions , publisher =. doi:10.1039/B606877D , abstract =

    work page doi:10.1039/b606877d

  29. [31]

    International conference on machine learning , pages=

    E (n) equivariant graph neural networks , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  30. [32]

    arXiv preprint arXiv:2511.07087 , year=

    Direct Molecular Polarizability Prediction with SO (3) Equivariant Local Frame GNNs , author=. arXiv preprint arXiv:2511.07087 , year=

    arXiv

  31. [33]

    arXiv preprint arXiv:2405.15389 , year=

    Beyond canonicalization: How tensorial messages improve equivariant message passing , author=. arXiv preprint arXiv:2405.15389 , year=

    arXiv

  32. [34]

    Journal of Chemical Theory and Computation , volume=

    Kernel-based minimal distributed charges: A conformationally dependent ESP-model for molecular simulations , author=. Journal of Chemical Theory and Computation , volume=. 2024 , publisher=

    2024

  33. [35]

    Journal of Chemical Theory and Computation , volume=

    Molecular dynamics with conformationally dependent, distributed charges , author=. Journal of Chemical Theory and Computation , volume=. 2022 , publisher=

    2022

  34. [36]

    Journal of chemical theory and computation , volume=

    Polarizable multipolar molecular dynamics using distributed point charges , author=. Journal of chemical theory and computation , volume=. 2020 , publisher=

    2020

  35. [37]

    , title =

    Stone, Anthony J. , title =. Journal of Chemical Theory and Computation , year =

  36. [38]

    Physical Chemistry Chemical Physics , year =

    Pracht, Philipp and Bohle, Fabian and Grimme, Stefan , title =. Physical Chemistry Chemical Physics , year =

  37. [39]

    Journal of Chemical Theory and Computation , year =

    Bannwarth, Christoph and Ehlert, Sebastian and Grimme, Stefan , title =. Journal of Chemical Theory and Computation , year =

  38. [40]

    Acta Crystallographica Section A , year =

    Kabsch, Wolfgang , title =. Acta Crystallographica Section A , year =

  39. [41]

    Chaudhry, Imran and Bronson, Mark J. Jr. and Jensen, Lasse , month = jul, year =. -. Journal of Chemical Theory and Computation , publisher =. doi:10.1021/acs.jctc.5c00752 , number =

    work page doi:10.1021/acs.jctc.5c00752

  40. [42]

    Molecular Dynamics Simulations Based on Polarizable Models Show that Ion Permeation Interconverts between Different Mechanisms as a Function of Membrane Thickness , journal =

    Chen, Peiran and Vorobyov, Igor and Roux, Beno. Molecular Dynamics Simulations Based on Polarizable Models Show that Ion Permeation Interconverts between Different Mechanisms as a Function of Membrane Thickness , journal =. 2021 , volume =

    2021

  41. [43]

    Analyzing Dynamical Disorder for Charge Transport in Organic Semiconductors via Machine Learning , journal =

    Reiser, Patrick and Konrad, Manuel and Fediai, Artem and Le. Analyzing Dynamical Disorder for Charge Transport in Organic Semiconductors via Machine Learning , journal =. 2021 , volume =

    2021

  42. [44]

    and Papageorgiou, D

    Kaklamanis, K. and Papageorgiou, D. G. , title =. Computer Physics Communications , year =

  43. [45]

    and Fu, Yao-Tsung and Risko, Chad and Br

    Ryno, Sean M. and Fu, Yao-Tsung and Risko, Chad and Br. Polarization Energies at Organic--Organic Interfaces: Impact on the Charge Separation Barrier at Donor--Acceptor Interfaces in Organic Solar Cells , journal =. 2016 , volume =

    2016

  44. [46]

    The Journal of Physical Chemistry A , author =

    Molecular and. The Journal of Physical Chemistry A , author =. 1998 , keywords =. doi:10.1021/jp980221f , language =

    work page doi:10.1021/jp980221f 1998