pith. sign in

arxiv: 1907.08717 · v1 · pith:AFHTUAP5new · submitted 2019-07-19 · ⚛️ physics.chem-ph · physics.comp-ph

Interpolation and extrapolation of global potential energy surfaces for polyatomic systems by Gaussian processes with composite kernels

Jun Dai , Roman V. Krems This is my paper

Pith reviewed 2026-05-24 18:38 UTC · model grok-4.3

classification ⚛️ physics.chem-ph physics.comp-ph
keywords Gaussian process regressionpotential energy surfacescomposite kernelsinterpolationextrapolationH3O+Bayesian information criterionpolyatomic molecules
0
0 comments X

The pith

Gaussian process models with composite kernels build accurate global six-dimensional PES for H3O+ from 500 ab initio points.

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

The paper demonstrates that Gaussian process regression for potential energy surfaces can achieve higher accuracy without adding more energy points by using composite kernels whose complexity is increased iteratively. It shows that the Bayesian information criterion can automate selection of these kernels and that accuracy improves further by adjusting the distribution of training points. The approach produces a global six-dimensional PES for H3O+ covering 0 to 21,000 cm^{-1} with 65.8 cm^{-1} RMSE from only 500 random ab initio points and extrapolates the same range from 1500 points below 10,000 cm^{-1}. A sympathetic reader would care because ab initio calculations are the dominant cost in constructing PES, so reducing the number of required points makes higher-dimensional or larger-molecule surfaces more tractable.

Core claim

Gaussian process regression models of potential energy surfaces trained with composite kernels, selected via the Bayesian information criterion, achieve a root mean square error of 65.8 cm^{-1} for the six-dimensional PES of H3O+ over 0 to 21,000 cm^{-1} using 500 ab initio points. The models can also extrapolate the PES to 21,000 cm^{-1} from 1500 points below 10,000 cm^{-1}. The accuracy is maximized by iteratively increasing kernel complexity and optimizing the distribution of training points.

What carries the argument

Composite kernels for Gaussian process regression, formed by combining multiple kernel functions to capture complex features of the potential energy surface.

If this is right

  • GP models with composite kernels produce global PES with high accuracy from small numbers of ab initio points.
  • The Bayesian information criterion automates selection of optimal kernel compositions for PES modeling.
  • Varying the distribution of training points further improves model accuracy for a fixed number of points.
  • GP models with composite kernels enable physical extrapolation of PES to higher energy regions from low-energy training data.

Where Pith is reading between the lines

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

  • The method could reduce the number of expensive ab initio calculations needed for PES of larger polyatomic systems.
  • Extrapolation performance could be tested on other molecular systems to establish how far the low-to-high energy transfer generalizes.
  • Pairing composite-kernel selection with active learning for point placement might lower the required training set size even further.

Load-bearing premise

Kernel compositions chosen by the Bayesian information criterion generalize to the full PES without overfitting to the particular distribution of the selected training points.

What would settle it

An independent test set of points above 10,000 cm^{-1} yields RMSE substantially larger than 65.8 cm^{-1} for a composite-kernel model trained only on points below 10,000 cm^{-1}.

Figures

Figures reproduced from arXiv: 1907.08717 by Jun Dai, Roman V. Krems.

Figure 1
Figure 1. Figure 1: FIG. 1: Schematic diagram of the composite kernel construction method. At each iteration, the [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2: RMSE for PES interpolation models with different kernels at kernel complexity level [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3: Upper panel: The dependence of the RMSE for the GP interpolation models of global [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4: Comparison of the GP prediction (solid curves) with the original potential energy points [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5: Comparison of the GP prediction with the original potential energy points for OH [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6: RMSE of the interpolation/extrapolation models trained by a fixed distribution of 1500 [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: Note that the energy points in Figure 5 include all [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7: Schematic depiction of the variables used to describe the OH [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8: Six different cuts of the potential energy surface for [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
read the original abstract

Gaussian process regression has recently emerged as a powerful, system-agnostic tool for building global potential energy surfaces (PES) of polyatomic molecules. While the accuracy of GP models of PES increases with the number of potential energy points, so does the numerical difficulty of training and evaluating GP models. Here, we demonstrate an approach to improve the accuracy of global PES without increasing the number of energy points. The present work reports four important results. First, we show that the selection of the best kernel function for GP models of PES can be automated using the Bayesian information criterion as a model selection metric. Second, we demonstrate that GP models of PES trained by a small number of energy points can be significantly improved by iteratively increasing the complexity of GP kernels. The composite kernels thus obtained maximize the accuracy of GP models for a given distribution of potential energy points. Third, we show that the accuracy of the GP models of PES with composite kernels can be further improved by varying the training point distributions. Fourth, we show that GP models with composite kernels can be used for physical extrapolation of PES. We illustrate the approach by constructing the six-dimensional PES for H$_3$O$^+$. For the interpolation problem, we show that this algorithm produces a global six-dimensional PES for H$_3$O$^+$ in the energy range between zero and $21,000$ cm$^{-1}$ with the root mean square error $65.8$ cm$^{-1}$ using only 500 randomly selected {\it ab initio} points as input. To illustrate extrapolation, we produce the PES at high energies using the energy points at low energies as input. We show that one can obtain an accurate global fit of the PES extending to $21,000$ cm$^{-1}$ based on 1500 potential energy points at energies below $10,000$ cm$^{-1}$.

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

3 major / 0 minor

Summary. The manuscript presents a Gaussian process regression approach using composite kernels selected by the Bayesian information criterion (BIC) to construct global potential energy surfaces (PES) for polyatomic systems. It reports four results: automated kernel selection, iterative complexity increase for better accuracy, improvement by varying training distributions, and physical extrapolation. For H3O+, it achieves RMSE of 65.8 cm^{-1} for interpolation up to 21,000 cm^{-1} with 500 ab initio points, and demonstrates extrapolation to the same range using 1500 low-energy points below 10,000 cm^{-1}.

Significance. If validated with proper hold-out testing, the method could reduce the number of ab initio points required for accurate global PES while automating kernel composition via BIC, offering a practical advance for computational chemistry of polyatomics. The concrete numerical demonstration on a six-dimensional H3O+ surface and the focus on extrapolation are strengths, though the absence of baseline comparisons and error estimates weakens immediate applicability.

major comments (3)
  1. [Abstract] Abstract: The reported interpolation RMSE of 65.8 cm^{-1} (500 randomly selected points) is presented without error bars, cross-validation statistics, or direct comparison to non-composite kernels, so it is impossible to determine whether the composite-kernel improvement is statistically meaningful or robust to point-distribution variation.
  2. [Abstract] Abstract (extrapolation paragraph): The claim that 1500 points below 10,000 cm^{-1} suffice for an accurate global fit to 21,000 cm^{-1} provides no RMSE, hold-out error, or other quantitative metric on the high-energy regime; BIC selection occurs exclusively on the low-energy training distribution, leaving the generalization assumption untested.
  3. [Abstract] Abstract (four results): No quantitative evidence is supplied for the second and third results (iterative kernel-complexity increase and varying training distributions), so the central assertion that composite kernels maximize accuracy for a given point set rests on the single interpolation number without supporting ablation data.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each of the major comments below and indicate the revisions we will make to improve the clarity and completeness of the abstract and supporting evidence.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The reported interpolation RMSE of 65.8 cm^{-1} (500 randomly selected points) is presented without error bars, cross-validation statistics, or direct comparison to non-composite kernels, so it is impossible to determine whether the composite-kernel improvement is statistically meaningful or robust to point-distribution variation.

    Authors: The abstract provides a summary of the key numerical result. The full manuscript describes the use of BIC for automated kernel selection and includes comparisons through the iterative process. To address the referee's concern, we will revise the abstract to mention that the reported RMSE is obtained after cross-validation and that composite kernels were selected over simpler alternatives based on BIC scores. revision: yes

  2. Referee: [Abstract] Abstract (extrapolation paragraph): The claim that 1500 points below 10,000 cm^{-1} suffice for an accurate global fit to 21,000 cm^{-1} provides no RMSE, hold-out error, or other quantitative metric on the high-energy regime; BIC selection occurs exclusively on the low-energy training distribution, leaving the generalization assumption untested.

    Authors: We agree that a specific quantitative metric for the extrapolation performance would strengthen the abstract. The manuscript shows the extrapolation by training exclusively on low-energy data and evaluating the model on the full range up to 21,000 cm^{-1}. In the revised version, we will include the RMSE achieved in the high-energy regime to quantify the extrapolation accuracy. Regarding BIC selection, it is performed on the training set as is standard, and the hold-out on high energies tests the generalization. revision: yes

  3. Referee: [Abstract] Abstract (four results): No quantitative evidence is supplied for the second and third results (iterative kernel-complexity increase and varying training distributions), so the central assertion that composite kernels maximize accuracy for a given point set rests on the single interpolation number without supporting ablation data.

    Authors: The manuscript provides demonstrations and supporting data for the iterative kernel complexity increase and the effect of training distributions in dedicated sections, including accuracy improvements shown in figures. The abstract condenses these into qualitative statements. We will update the abstract to include brief quantitative indications of the improvements from these steps, such as the reduction in error from iterative composition. revision: yes

Circularity Check

0 steps flagged

No significant circularity; results benchmarked against external ab initio data

full rationale

The paper trains GP models (with BIC-selected composite kernels) on subsets of ab initio points for H3O+ and reports RMSE on other ab initio points for interpolation (500 training points) or on high-energy points for extrapolation (1500 low-energy training points). These are standard held-out evaluations against independent quantum-chemistry calculations, not reductions to the training inputs by construction. No self-citations, uniqueness theorems, or ansatzes are invoked as load-bearing steps in the provided text. The derivation chain consists of data-driven model selection followed by direct numerical comparison to external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The approach rests on standard assumptions of Gaussian process regression and the appropriateness of BIC for kernel selection in this domain; no new entities are postulated.

free parameters (1)
  • kernel hyperparameters
    Standard GP parameters fitted to the ab initio energy points; their number increases with composite kernel complexity.
axioms (2)
  • domain assumption PES are sufficiently smooth for GP regression with standard kernels to be applicable
    Invoked implicitly when applying GP to molecular energies.
  • domain assumption BIC is a reliable metric for selecting kernels that generalize beyond the training set
    Used as the model-selection criterion without further justification in the abstract.

pith-pipeline@v0.9.0 · 5874 in / 1382 out tokens · 25439 ms · 2026-05-24T18:38:51.429771+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

47 extracted references · 47 canonical work pages · 1 internal anchor

  1. [1]

    These points thus sample the regions both within and outside the range of coordinates of the training points

    Specifically, Figure 5 includes 26383 points above 5000 cm−1; 22386 points above 7000 cm−1; 16252 points above 10000 cm−1; and 6950 points above 15000 cm−1. These points thus sample the regions both within and outside the range of coordinates of the training points. The RMSE of the PES thus obtained are shown in Figure 6. As is clear from the numerical val...

    work page

  2. [2]

    There is thus absolutely no information about this part of the 14 configuration space in the training set of energy points

    The set of ab initio points does not describe this part of the configuration space. There is thus absolutely no information about this part of the 14 configuration space in the training set of energy points. The different curves in Figure 8 correspond to predictions of five different GP models obtained, as before, by GP regression of the training points at ene...

    work page 2000

  3. [3]

    B. J. Braams and J. M. Bowman, Permutationally invariant potential energy surfaces in high dimensionality,Int. Rev. Phys. Chem.28, 577 (2009)

    work page 2009

  4. [4]

    J. N. Murrell, S. Carter, S. C. Farantos, P. Huxley, and A. J. C. Varandas,Molecular Potential Energy Functions(Wiley, Chichester, England, 1984)

    work page 1984

  5. [5]

    Hollebeek, T.-S

    T. Hollebeek, T.-S. Ho, and H. Rabitz, Constructing multidimensional molecular potential energy surfaces fromab initiodata, Annu. Rev. Phys. Chem.50, 537 (1999)

    work page 1999

  6. [6]

    M. A. Collins, Molecular potential-energy surfaces for chemical reaction dynamics, Theor. Chem. Acc.108, 313 (2002)

    work page 2002

  7. [7]

    C. M. Handley and P. L. A. Popelier, Potential Energy Surfaces Fitted by Artificial Neural Networks, J. Phys. Chem.A 114, 3371 (2010)

    work page 2010

  8. [8]

    R. V. Krems, Bayesian Machine Learning for Quantum Molecular Dynamics,Phys. Chem. Chem. Phys. 21, 13392 (2019)

    work page 2019

  9. [9]

    Behler, Perspective: Machine learning potentials for atomistic simulations, J

    J. Behler, Perspective: Machine learning potentials for atomistic simulations, J. Chem. Phys.145, 170901 (2016)

    work page 2016

  10. [10]

    Manzhos and T

    S. Manzhos and T. Carrington, Jr., A random-sampling high dimensional model representation neural network for building potential energy surfacesJ. Chem. Phys.125, 084109 (2006)

    work page 2006

  11. [11]

    Manzhos, X

    S. Manzhos, X. Wang, R. Dawes, and T. Carrington, Jr., A nested molecule-independent neural network approach for high-quality potential fits,J. Phys. Chem.A 110, 5295 (2006)

    work page 2006

  12. [12]

    J Behler and M Parrinello, Generalized neural-network representation of high-dimensional potential-energy surfaces,Phys. Rev. Lett.98, 146401 (2007)

    work page 2007

  13. [13]

    Behler, Neural network potential-energy surfaces in chemistry: a tool for large-scale simu- lations, Phys

    J. Behler, Neural network potential-energy surfaces in chemistry: a tool for large-scale simu- lations, Phys. Chem. Chem. Phys.13, 17930 (2011)

    work page 2011

  14. [14]

    Behler, Constructing high-dimensional neural network potentials: A tutorial review,Int

    J. Behler, Constructing high-dimensional neural network potentials: A tutorial review,Int. J. Quant. Chem. 115, 1032 (2015)

    work page 2015

  15. [15]

    Pradesh and A

    E. Pradesh and A. Brown, A ground state potential energy surface for HONO based on a neural network with exponential fitting functions,Phys. Chem. Chem. Phys.19, 22272 (2017)

    work page 2017

  16. [16]

    Leclerc and T

    A. Leclerc and T. Carrington, Jr., Calculating vibrational spectra with sum of product basis functions without storing full-dimensional vectors or matrices,J. Chem. Phys. 140, 174111 (2014). 19

    work page 2014

  17. [17]

    Manzhos, R

    S. Manzhos, R. Dawes, and T. Carrington, Neural network-based approaches for building high dimensional and quantum dynamics-friendly potential energy surfaces,Int. J. Quant. Chem. 115, 1012 (2015)

    work page 2015

  18. [18]

    J. Chen, X. Xu, X. Xu, and D. H. Zhang, A global potential energy surface for the H2 + OH ↔ H2O + H reaction using neural networks,J. Chem. Phys.138, 154301 (2013)

    work page 2013

  19. [19]

    Q. Liu, X. Zhou, L. Zhou, Y. Zhang, X. Luo, H. Guo, and B. Jiang, Constructing High- Dimensional Neural Network Potential Energy Surfaces for Gas-Surface Scattering and Reac- tions, J. Phys. Chem.C 122, 1761 (2018)

    work page 2018

  20. [20]

    C. M. Handley, G. I. Hawe, D. B. Kellab and P. L. A. Popelier, Optimal construction of a fast and accurate polarisable water potential based on multipole moments trained by machine learning, Phys. Chem. Chem. Phys.11, 6365 (2009)

    work page 2009

  21. [21]

    A. P. Bartók, M. C. Payne, R. Kondor, and G. Csányi, Gaussian Approximation Potentials: The Accuracy of Quantum Mechanics, without the Electrons,Phys. Rev. Lett. 104, 136403 (2010)

    work page 2010

  22. [22]

    A. P. Bartók and G. Csányi, Gaussian approximation potentials: A brief tutorial introduction, Int. J. Quant. Chem.115, 1051 (2015)

    work page 2015

  23. [23]

    Cui and R

    J. Cui and R. V. Krems, Efficient non-parametric fitting of potential energy surfaces for poly- atomic molecules with Gaussian processes,J. Phys. B: At. Mol. Opt. Phys.49, 224001 (2016)

    work page 2016

  24. [24]

    P. O. Dral, A. Owens, S. N. Yurchenko, and W. Thiel, Structure-based sampling and self- correcting machine learning for accurate calculations of potential energy surfaces and vibra- tional levels,J. Chem. Phys.146, 244108 (2017)

    work page 2017

  25. [25]

    B. Kolb, P. Marshall, B. Zhao, B. Jiang, and Hua Guo, Representing Global Reactive Potential Energy Surfaces Using Gaussian Processes,J. Phys. Chem.A 121, 2552 (2017)

    work page 2017

  26. [26]

    Kamath, R

    A. Kamath, R. A. Vargas-Hernandez, R. V. Krems, T. Carrington Jr., and S. Manzhos, Neural networks vs Gaussian process regression for representing potential energy surfaces: A compar- ative study of fit quality and vibrational spectrum accuracy,J. Chem. Phys. 148, 241702 (2018)

    work page 2018

  27. [27]

    Schmitz and O

    G. Schmitz and O. Christiansen, Gaussian process regression to accelerate geometry optimiza- tions relying on numerical differentiationJ. Chem. Phys.148, 241704 (2018)

    work page 2018

  28. [28]

    Y. Guan, S. Yang, and D. H. Zhang, Construction of reactive potential energy surfaces with Gaussian process regression: active data selection,Mol. Phys. 116, 823 (2018). 20

    work page 2018

  29. [29]

    Laude, D

    G. Laude, D. Calderini, D. P. Tew, and J. O. Richardson,ab initioinstanton rate theory made efficient using Gaussian process regression,Faraday Discuss.212, 237 (2018)

    work page 2018

  30. [30]

    Y. Guan, S. Yang, and D. H. Zhang, Application of Clustering Algorithms to Partitioning Configuration Space in Fitting Reactive Potential Energy Surfaces,J. Phys. Chem. A 122, 3140 (2018)

    work page 2018

  31. [31]

    A. E. Wiens, A. V. Copan, H. F. Schaefer, Multi-Fidelity Gaussian Process Modeling for Chemical Energy Surfaces,Chem. Phys. Lett.X, in press (2019)

    work page 2019

  32. [32]

    C. Qu, Q. Yu, B. L. Van Hoozen Jr, J. M. Bowman, and R. A. Vargas-Hernandez, Assessing Gaussian Process Regression and Permutationally Invariant Polynomial Approaches To Rep- resent High-Dimensional Potential Energy Surfaces,J. Chem. Theor. Comp.14, 3381 (2018)

    work page 2018

  33. [33]

    T. S. Ho and H. Rabitz, A general method for constructing multidimensional molecular po- tential energy surfaces fromab initiocalculations, J. Chem. Phys.104, 2584 (1996)

    work page 1996

  34. [34]

    Hollebeek, T

    T. Hollebeek, T. S. Ho, and H. Rabitz, A fast algorithm for evaluating multidimensional potential energy surfaces,J. Chem. Phys.106, 7223 (1997)

    work page 1997

  35. [35]

    T. S. Ho and H. Rabitz, Reproducing kernel Hilbert space interpolation methods as a paradigm of high dimensional model representations: Application to multidimensional potential energy surface construction,J. Chem. Phys.119, 6433 (2003)

    work page 2003

  36. [36]

    Christianen, T

    A. Christianen, T. Karman, R. A. Vargas-Hernandez, G. C. Groenenboom and R. V. Krems, Six-dimensional potential energy surface for NaK-NaK collisions: Gaussian process represen- tation with correct asymptotic form,J. Chem. Phys.150, 064106 (2019)

    work page 2019

  37. [37]

    R.V.Krems, Molecules in electromagnetic fields: from ultracold physics to controlled chemistry, Wiley (2018)

    work page 2018

  38. [38]

    R.Vargas-Hernandez, Y. Guan, D. H. Zhang, and R. V. Krems, Bayesian optimization for the inverse scattering problem in quantum reaction dynamics,New J. Phys.(Fast Track Commu- nication) 21, 022001 (2019)

    work page 2019

  39. [39]

    D. K. Duvenaud, H. Nickisch, and C. E. Rasmussen, Additive Gaussian Processes,Adv. Neur. Inf. Proc. Sys.24, 226 (2011)

    work page 2011

  40. [40]

    D. K. Duvenaud, J. Lloyd, R. Grosse, J. B. Tenenbaum, and Z. Ghahramani, Structure Dis- covery in Nonparametric Regression through Compositional Kernel Search,Proceedings of the 30th International Conference on Machine Learning Research28, 1166 (2013)

    work page 2013

  41. [41]

    R.Vargas-Hernandez, J.Sous, M.Berciu, andR.V.Krems, Extrapolatingquantumobservables 21 with machine learning: Inferring multiple phase transitions from properties of a single phase, Phys. Rev. Lett.121, 255702 (2018)

    work page 2018

  42. [42]

    M.ab initio potential for H3O+→ H+ + H2O: A step to a many-body representation of the hydrated proton?J

    Yu, Q.; Bowman, J. M.ab initio potential for H3O+→ H+ + H2O: A step to a many-body representation of the hydrated proton?J. Chem. Theor. Comp.12, 5284 (2016)

    work page 2016

  43. [43]

    C. E. Rasmussen, and C. K. I. Williams,Gaussian Processes for Machine Learning(The MIT Press, Cambridge, 2006)

    work page 2006

  44. [44]

    Schwarz, Estimating the dimension of a model,The Annals of Statistics6(2), 461 (1978)

    G. Schwarz, Estimating the dimension of a model,The Annals of Statistics6(2), 461 (1978)

    work page 1978

  45. [45]

    Sugisawa, R

    H. Sugisawa, R. Vargas-Hernandez and R. V. Krems, to be published (2019)

    work page 2019

  46. [46]

    Hensman, N

    J. Hensman, N. Fusi, N. D. Lawrence,Gaussian processes for big data, Proceedings of the Twenty-Ninth Conference on Uncertainty in Artificial Intelligence (UAI2013), Report number: UAI-P-2013-PG-282-290 (2013)

    work page 2013

  47. [47]

    When Gaussian Process Meets Big Data: A Review of Scalable GPs

    H. Liu, Y.-S. Ong, X. Shen, and J. Cai,When Gaussian Process Meets Big Data: A Review of Scalable GPs, arXiv:1807.01065. 22

    work page internal anchor Pith review Pith/arXiv arXiv