Quantifying nuclear quantum effects and functional-driven errors in the overestimated dielectric constant of machine-learned water

Machine-learned interatomic potentials with explicit long-range electrostatics, trained on hybrid density functional theory (DFT) data at the SCAN0 level and combined with a neural network that predicts maximally localized Wannier centers, predict a static dielectric constant of about 102 for liquid water with classical nuclei at 330 K and 1 bar, substantially above the experimental value (about 78 at ambient conditions, and below 70 at 330 K). The source study attributes this overestimation to two unquantified sources: the absence of nuclear quantum effects and residual exchange-correlation errors of the training functional. The open problem is a quantified account of this discrepancy: compute the dielectric constant of water within the long-range machine-learning framework including nuclear quantum effects via path-integral molecular dynamics, measure how the molecular dipole moment, Kirkwood correlation factor and correlation length, dielectric relaxation times, and the dielectric constant itself change; isolate the residual functional-driven error; and determine whether the resulting first-principles prediction, with uncertainties, agrees with experiment.

Background

The static dielectric constant $\varepsilon$ of liquid water (about 78 at room temperature) measures how strongly water screens electric fields and underlies its behavior as the universal solvent, yet predicting it from first principles has been difficult for decades. In the linear-response (dipole fluctuation) approach, $\varepsilon$ is obtained from the equilibrium fluctuations of the total dipole moment $\mathbf{M}$ of a simulation cell, or equivalently from the Kirkwood-Frohlich formalism, in which $\varepsilon$ is built from the molecular dipole moment $\mu$ and the Kirkwood correlation factor $G_K$ — the integral of the dipole-dipole correlation function, which measures how far orientational correlations extend through the hydrogen-bond network (about 16 Angstrom in water, requiring cells of roughly a thousand molecules or more). A subtlety is that $\mathbf{M}$-fluctuations depend on the electric boundary condition imposed on the periodically repeated cell: a metallic boundary condition (surrounding medium of permittivity $\varepsilon'=\infty$, the closed-circuit or tinfoil setting), an insulating boundary condition ($\varepsilon'=0$, the open-circuit setting of constant electric displacement, studied systematically by Zhang and Sprik's constant-displacement method), and the Kirkwood-Frohlich boundary condition ($\varepsilon'=\varepsilon$, which requires knowing the answer in advance). The three give consistent $\varepsilon$ only when the long-range dipolar correlations and the reaction-field corrections — the finite-size corrections to the fluctuation formula induced by the permittivity $\varepsilon'$ of the surrounding medium — are treated correctly.

Ab initio molecular dynamics (AIMD) computes forces from density functional theory (DFT), but converged dipole fluctuations demand tens of nanoseconds and thousands of molecules — far beyond direct AIMD, as recognized since the pioneering ab initio study of the dipolar correlations behind water's $\varepsilon$ by Sharma, Resta, and Car. Maximally localized Wannier functions (MLWFs) are localized orbitals obtained from the DFT eigenstates by a unitary transformation; their centers play the role of effective negative charge centers, so the molecular dipole is the sum of ionic and MLWF-center contributions. Two deep-neural-network components make DFT-quality large-scale sampling tractable: the Deep Potential Long-Range (DPLR) model, a machine-learned potential that adds explicit long-range Coulomb interactions between Gaussian charges at ionic and (predicted) electronic sites, and the Deep Wannier model, which predicts MLWF centers from nuclear configurations so that dipoles come out of the same framework. Related efforts include the demonstration that short-range electrostatics can recover dielectric response in homogeneous phases (Cox), and self-consistent schemes for embedding long-range electrostatics in neural network potentials (Gao and Remsing).

The source study (Cai, Zhang, and Wu, J. Chem. Phys. 163, 184503 (2025)) trained DPLR and Deep Wannier models on SCAN0 hybrid DFT data (SCAN is a meta-GGA exchange-correlation functional; SCAN0 mixes a fraction of exact exchange — 10% in this work — into it to reduce self-interaction error) and simulated 4096 water molecules for 20 ns at 330 K and 1 bar. It consistently obtains $\varepsilon \approx 102$ from both the dipole fluctuation and Kirkwood correlation methods, under metallic, Kirkwood-Frohlich, and combined metallic-insulating boundary conditions, with molecular dipole $\mu \approx 2.97$ Debye and Kirkwood factor $G_K \approx 2.4$. Nuclear quantum effects (NQEs) — the quantum delocalization of the light hydrogen nuclei, computable with path-integral molecular dynamics (PIMD), in which each quantum nucleus is mapped onto a classical ring polymer of beads — were absent, since the trajectories used classical nuclei. The paper attributes the overestimation relative to experiment to this absence of NQEs and to residual SCAN0 exchange-correlation error, without quantifying either. The attribution is nontrivial: the same group's earlier SCAN0 study with path-integral deep-potential simulations found that NQEs raise $\varepsilon$ (from 76.06 to 83.57 at 330 K), a SCAN-trained neural-network study reported experimental agreement with classical nuclei, and the experimental dielectric constants of light and heavy water differ by only about 0.4 — so what moves the prediction by tens of percent, and in which direction, is genuinely unsettled.

Problem Statement

Within the long-range machine-learning framework of the source study — a Deep Potential Long-Range (DPLR) potential for the forces plus a Deep Wannier network for the Wannier-center dipoles, both trained on SCAN0 hybrid DFT data, which predicts a static dielectric constant $\varepsilon \approx 102$ for liquid water with classical nuclei (NPT ensemble, i.e. simulation at constant pressure and temperature, 1 bar, 330 K, 4096 molecules, 20-nanosecond trajectories), substantially above the experimental value (about 78 at ambient conditions, below 70 at 330 K) — determine, with quantified statistical uncertainty and at DFT-level accuracy, the origin of the overestimation. Specifically:

1. (Nuclear quantum contribution) Compute $\varepsilon$ with nuclear quantum effects included via path-integral molecular dynamics within the source DPLR-plus-Deep-Wannier framework (the same class of machine-learned models as the source study), and quantify how the inclusion of NQEs changes the molecular dipole moment $\mu$, the Kirkwood correlation factor $G_K$ and correlation length, the transverse and longitudinal dielectric relaxation times, and $\varepsilon$ itself. The calculation must use the same electric boundary-condition methodology as the source framework (metallic, insulating, Kirkwood-Frohlich, or the combined metallic-insulating reconstruction, with the corresponding reaction-field corrections), simulation cells exceeding the roughly 16-Angstrom dipolar correlation length, and trajectories converged for the dipole fluctuations (or an equivalently validated variance-reduction scheme), with demonstrated convergence in path-integral bead number.

2. (Functional contribution) Isolate the residual error attributable to the SCAN0 exchange-correlation approximation, so that the total overestimation is decomposed into a nuclear-quantum part and a functional-driven part.

3. (Comparison with experiment) State whether the resulting prediction, including NQEs and the best available reference level, agrees with the experimental dielectric constant at the simulated state point within combined uncertainties. Conflicting same-level predictions already exist in the published literature (classical SCAN0 deep-potential $\varepsilon = 76.06$ and path-integral $\varepsilon = 83.57$ at 330 K, obtained with a short-range deep-potential model class distinct from the source's long-range framework and therefore not supplying the source-framework NQE contribution; the experimental agreement reported by SCAN-trained neural-network quantum molecular dynamics; and the classical $\varepsilon = 87 \pm 1$ of a CCSD(T)-level range-separated water model, where CCSD(T) is a high-accuracy coupled-cluster quantum-chemistry method); these values provide context for interpreting the decomposition, but reproducing or reconciling them is not part of the required answer.

References

1. K. Cai, C. Zhang, and X. Wu, "Simulations of dielectric permittivity of water by machine learned potentials with long-range Coulombic interactions", The Journal of Chemical Physics 163, 184503 (2025). DOI: 10.1063/5.0294554, https://doi.org/10.1063/5.0294554 (preprint arXiv:2508.04628, https://arxiv.org/abs/2508.04628)
2. C. Zhang, F. Tang, M. Chen, J. Xu, L. Zhang, D. Y. Qiu, J. P. Perdew, M. L. Klein, and X. Wu, "Modeling Liquid Water by Climbing up Jacob's Ladder in Density Functional Theory Facilitated by Using Deep Neural Network Potentials", The Journal of Physical Chemistry B 125, 11444-11456 (2021). DOI: 10.1021/acs.jpcb.1c03884, https://doi.org/10.1021/acs.jpcb.1c03884 (preprint arXiv:2104.14410, https://arxiv.org/abs/2104.14410)
3. A. Krishnamoorthy, K. Nomura, N. Baradwaj, K. Shimamura, P. Rajak, A. Mishra, S. Fukushima, F. Shimojo, R. Kalia, A. Nakano, and P. Vashishta, "Dielectric Constant of Liquid Water Determined with Neural Network Quantum Molecular Dynamics", Physical Review Letters 126, 216403 (2021). DOI: 10.1103/PhysRevLett.126.216403, https://doi.org/10.1103/PhysRevLett.126.216403
4. C. Zhang and M. Sprik, "Computing the dielectric constant of liquid water at constant dielectric displacement", Physical Review B 93, 144201 (2016). DOI: 10.1103/PhysRevB.93.144201, https://doi.org/10.1103/PhysRevB.93.144201
5. M. Sharma, R. Resta, and R. Car, "Dipolar Correlations and the Dielectric Permittivity of Water", Physical Review Letters 98, 247401 (2007). DOI: 10.1103/PhysRevLett.98.247401, https://doi.org/10.1103/PhysRevLett.98.247401
6. S. J. Cox, "Dielectric response with short-ranged electrostatics", Proceedings of the National Academy of Sciences 117, 19746-19752 (2020). DOI: 10.1073/pnas.2005847117, https://doi.org/10.1073/pnas.2005847117
7. A. Gao and R. C. Remsing, "Self-consistent determination of long-range electrostatics in neural network potentials", Nature Communications 13, 1572 (2022). DOI: 10.1038/s41467-022-29243-2, https://doi.org/10.1038/s41467-022-29243-2
8. L. Zhang, M. Chen, X. Wu, H. Wang, W. E, and R. Car, "Deep neural network for the dielectric response of insulators", Physical Review B 102, 041121 (2020). DOI: 10.1103/PhysRevB.102.041121, https://doi.org/10.1103/PhysRevB.102.041121
9. L. Zhang, H. Wang, M. C. Muniz, A. Z. Panagiotopoulos, R. Car, and W. E, "A deep potential model with long-range electrostatic interactions", The Journal of Chemical Physics 156, 124107 (2022). DOI: 10.1063/5.0083669, https://doi.org/10.1063/5.0083669
10. Y. Li, A. Gomez, K. Cai, C. Zhang, L. Fu, W. Jia, Y. M. Y. Feldman, O. Blumer, J. Higer, B. Hirshberg, S. Xu, A. Kohlmeyer, and R. Car, "fix pimd/langevin: An efficient implementation of path integral molecular dynamics in LAMMPS", The Journal of Chemical Physics 164, 144119 (2026). DOI: 10.1063/5.0328429, https://doi.org/10.1063/5.0328429 (preprint arXiv:2602.13553, https://arxiv.org/abs/2602.13553)
11. Q. Gao, J. Chen, and K. Yu, "Refinement and performance benchmark for range-separated water force field", The Journal of Chemical Physics 164, 154109 (2026). DOI: 10.1063/5.0324748, https://doi.org/10.1063/5.0324748 (preprint arXiv:2601.18416, https://arxiv.org/abs/2601.18416)