A deep potential model with long-range electrostatic interactions

Linfeng Zhang, Handong Wang, Maria Carolina Muniz, Athanassios Z. Panagiotopoulos, Roberto Car, E WeinanThe Journal of Chemical Physics 2022 · DOI

TL;DR

This paper introduces the Deep Potential Long-Range (DPLR) model, an extension of existing machine learning potentials for molecular simulations. DPLR accurately captures long-range electrostatic interactions, which are crucial for describing many material properties but are missing in standard models, leading to improved accuracy and predictive power, especially for larger systems.

What Problem Does This Paper Address?

Machine learning (ML) models for predicting molecular potential energy surfaces (PESs) offer a powerful way to bridge the gap between the accuracy of quantum mechanics (like DFT) and the efficiency of empirical force fields. However, a significant limitation of current ML potentials, including the widely used Deep Potential (DP) model, is their reliance on short-range interactions. This means they struggle to accurately describe phenomena driven by the long-range Coulombic forces, such as polarization effects in insulators, the behavior of polar materials, and properties of extended systems. This deficiency leads to errors in predicting certain physical properties and limits the ability to extrapolate models to larger system sizes than those used during training.

Methods

The core innovation is the Deep Potential Long-Range (DPLR) model. It builds upon the existing Deep Potential (DP) framework by explicitly incorporating long-range electrostatic interactions. This is achieved by:

  • Approximating electrostatic interactions: The long-range electrostatic energy is modeled using spherical Gaussian charge distributions placed at ionic sites (nuclei + core electrons) and at "electronic sites."
  • Defining electronic sites: These electronic sites are determined by the centers of maximally localized Wannier distributions (MLWCs). The dependence of these MLWC positions on the local atomic environment is learned by a separate deep neural network, called the deep Wannier (DW) model.
  • Combining short and long-range contributions: The total PES in DPLR is the sum of the short-range interactions (as in the standard DP model) and the calculated long-range electrostatic energy.
  • Analytical forces and virial: The DPLR model is designed to yield smooth PESs, allowing for the analytical calculation of forces and virial, which is essential for stable molecular dynamics simulations.
  • Efficient calculation: The electrostatic energy is computed efficiently using fast algorithms like the particle-particle-particle-mesh (PPPM) method for Ewald sums.

Datasets

  • Training Data: The models (DP, DW, and DPLR) were trained using data generated by the Deep Potential GENerator (DP-GEN) protocol. This protocol iteratively expands the training dataset by sampling relevant thermodynamic spaces (bulk, surface, low-density configurations) using deep potential molecular dynamics (DPMD) and performing single-shot DFT calculations on selected configurations.
  • DFT Calculations: DFT calculations were performed using the Vienna Ab initio Simulation Package (VASP) with the PBE exchange-correlation functional.
  • Test Data: The models were evaluated on independent configurations not used in training, including bulk and surface configurations for water, and configurations for crystalline NaCl.

Key Insights and Ideas

  • The "Nearsightedness" Principle: The paper leverages the concept of electronic "nearsightedness," which suggests that the electronic properties of a material are primarily determined by the local atomic environment. This justifies modeling the dependence of Wannier centroids on a local atomic environment within a cutoff radius.
  • Wannier Centroids as Electronic Sites: Using Wannier centroids as the locations for electronic charge distributions provides a more physically grounded approach than simple atom-centered charges, especially for capturing polarization effects.
  • Optimal Spread Parameter: The study highlights the importance of an "optimal spread parameter" (β) for the Gaussian charge distributions. Too small a β makes DPLR reduce to the standard DP model, while too large a β leads to numerical issues due to highly localized charges. An optimal value was found to be around 0.4 Å⁻¹ for water, correlating with the physical extent of electron density.
  • Extrapolation Capability: A key advantage of DPLR is its ability to extrapolate the learned PES to larger system sizes, a limitation of standard short-range ML models. This is demonstrated by the size-dependent behavior of phonon dispersion curves in NaCl.

Key Results

  • Water Dimer: DPLR accurately reproduces the long-range Coulombic tail of the water dimer's potential energy curve, which is missed by the standard DP model.
  • Water-Slab Interaction: DPLR captures the attractive dipolar interactions between a water molecule and a liquid water slab, leading to a more accurate free energy profile compared to DP. The uncertainty in DPLR predictions was significantly smaller than that of DP when the molecule approached the slab.
  • NaCl Phonons: DPLR correctly captures the size dependence of phonon dispersion curves in NaCl, including the longitudinal-optical (LO) mode, which is crucial for polar materials. The standard DP model showed minimal size dependence, indicating its inability to capture these long-range effects.
  • Energy Drift: In molecular dynamics simulations of liquid water, DPLR showed a small energy drift (≈ 0.4 meV/H₂O per 100 ps) that was controllable by the accuracy of the Ewald sum calculation (mesh spacing). This drift was slightly larger than that of the DP model but still within acceptable limits.
  • Computational Cost: DPLR simulations are estimated to be approximately 5 times more expensive than standard DP calculations due to the added complexity of Ewald sums and backward propagation through both DP and DW networks.

Limitations and Future Work

  • Assumption of WC Uniqueness: The current DPLR model assumes that Wannier centroids are uniquely associated with specific atoms and do not split or recombine. This limitation restricts its applicability to chemical reactions involving electron transfer, such as proton transfer. The authors suggest relaxing this assumption for future generalizations.
  • Linear Response Approximation: The model assumes that long-range electrostatic effects beyond the DP cutoff radius originate a weak dependence on size, treatable within linear response theory. The authors acknowledge that in circumstances where this assumption doesn't hold, a self-consistent treatment of centroids under the influence of the long-range electrostatic field might be necessary, similar to approaches in Ref. [54].
  • Computational Cost: While DPLR offers improved accuracy, it comes with a higher computational cost compared to standard DP. The decision to use DPLR should be guided by the importance of long-range correlations for the specific application.
  • Future Work: The authors envision future work involving a more generalized DPLR model that can handle electron transfer reactions and exploring scenarios where the linear response approximation might not be sufficient.

Summary updated

Abstract

Machine learning models for the potential energy of multi-atomic systems, such as the deep potential (DP) model, make molecular simulations with the accuracy of quantum mechanical density functional theory possible at a cost only moderately higher than that of empirical force fields. However, the majority of these models lack explicit long-range interactions and fail to describe properties that derive from the Coulombic tail of the forces. To overcome this limitation, we extend the DP model by approximating the long-range electrostatic interaction between ions (nuclei + core electrons) and valence electrons with that of distributions of spherical Gaussian charges located at ionic and electronic sites. The latter are rigorously defined in terms of the centers of the maximally localized Wannier distributions, whose dependence on the local atomic environment is modeled accurately by a deep neural network. In the DP long-range (DPLR) model, the electrostatic energy of the Gaussian charge system is added to short-range interactions that are represented as in the standard DP model. The resulting potential energy surface is smooth and possesses analytical forces and virial. Missing effects in the standard DP scheme are recovered, improving on accuracy and predictive power. By including long-range electrostatics, DPLR correctly extrapolates to large systems the potential energy surface learned from quantum mechanical calculations on smaller systems. We illustrate the approach with three examples: the potential energy profile of the water dimer, the free energy of interaction of a water molecule with a liquid water slab, and the phonon dispersion curves of the NaCl crystal.
BibTeX
@article{zhang2022deep,
  author = {Linfeng Zhang and Handong Wang and Maria Carolina Muniz and Athanassios Z. Panagiotopoulos and Roberto Car and E Weinan},
  title = {{A deep potential model with long-range electrostatic interactions}},
  journal = {The Journal of Chemical Physics},
  year = {2022},
  doi = {10.1063/5.0083669},
}
Figure