Electrostatic interactions in atomistic and machine-learned potentials for polar materials

Lorenzo Monacelli, Nicola Marzari · 2024

Read Electrostatic interactions in atomistic and machine-learned potentials for polar materials by Lorenzo Monacelli, Nicola Marzari (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-01642.txt.

Introduction

Lorenzo Monacelli and Nicola Marzari wrote this paper in two thousand twenty-four. It adds the long-range electric forces that most atomistic potentials skip, in a form that plugs into the tools labs already use.

Neural and Gaussian potentials drive big simulations, yet they drop far-field charge effects that rule polar materials. The new model comes from first principles and needs no fresh such runs or retraining to attach. It rests on things that can be got the same way, like the dielectric tensor and Born charges. It then captures key marks of polar solids, such as the split of lengthwise and crosswise optical modes. Checks on barium titanate's cubic phase back the claims.

Catalog record

CREATORLorenzo Monacelli, Nicola Marzari
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE38 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-01642

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Electrostatic interactions in atomistic and machine-learned potentials for polar materials Lorenzo Monacelli, Nicola Marzari Abstract Long-range electrostatic interactions critically affect polar materials. However, state-of-the-art atomistic potentials, such as neural networks or Gaussian approximation potentials employed in large-scale simulations, often neglect the role of these long-range electrostatic interactions. This study introduces a novel model derived from first principles to evaluate the contribution of long-range electrostatic interactions to total energies, forces, and stresses. The model is designed to integrate seamlessly with existing short-range force fields without further first-principles calculations or retraining. The approach relies solely on physical observables, like the dielectric tensor and Born effective charges, that can be consistently calculated from first principles. We demonstrate that the model reproduces critical features, such as the LO-TO splitting and the long-wavelength phonon dispersions of polar materials, with benchmark results on the cubic phase of barium titanate ( BaTiO 3 \text{BaTiO}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}} ). Preprint: APS/123-QED I Introduction In polar materials, displacing an ion from equilibrium generates an electric dipole. The resulting electric field decays as the inverse of the cube of the distance and interacts with local dipoles generated by far-away ionic displacements. While these long-range electrostatic interactions are screened in metals, and their role becomes relevant only when accounting for ionic dynamics [ 1 , 2 ] , in the absence of electrons in the conduction band long-range electric fields survive down to zero frequency, thus playing a fundamental role in the phenomenology of insulators. This includes the splitting between longitudinal and transverse optical phonons at long wavelengths (LO-TO splitting) [ 3 , 4 , 5 , 6 ] , affecting in turn all thermodynamic properties related to phonon dispersions, such as thermal expansion, heat capacity, lattice thermal conductivity, and Raman and IR spectra. First-principles simulations include long-range electrostatic interactions [ 7 , 4 ] when considering perturbations of finite wavevector, where the phonons can be evaluated either by finite-differences [ 8 ] or density-functional perturbation theory (DFPT) [ 4 ] . At \Gamma ( \bm{q}=0 ), the contribution of electrostatic interaction is nonanalytic, as \bm{q}=0 periodic displacements produce macroscopic polarizations resulting in a dynamical matrix for phonons dependent on the direction of the \bm{q} vector as it approaches \Gamma . This nonanalytic behavior of lattice force constants near \Gamma constitutes a challenge in modeling efficient force fields for polar materials, as it originates from the cubic power-law decay of electrostatic forces and can not be captured if the interaction is truncated over a finite distance. Thanks to the advances in machine learning technologies, many excellent tools to fit atomistic potentials from small sets of high quality ab initio calculations have emerged [ 9 , 10 , 11 , 12 , 13 , 14 ] , paving the way to the simulation of large-scale materials with hundreds of millions of atoms with an accuracy close to first principles methods [ 15 , 16 ] . However, the most commonly employed machine-learning potentials neglect the long-range electrostatic interactions, resulting in a smooth \bm{q} dependence of the atomic force constant matrix around \Gamma . This is because the force field parametrizations account only for local atomic environments, and the first principles training data are evaluated on periodic boundary conditions that lack long-range contributions. The question arises of what should be fitted and how electrostatics can be recovered. Several methods have been developed to overcome this issue. Some approaches, the so-called second and third-generation machine-learning potential [ 17 ] , split the energy expression into a short and long-range part, explicitly accounting for electrostatic interactions between point charges dependent on the local nuclear environment [ 18 , 19 , 20 , 21 , 22 , 23 ] . These models suffer from the need for extra training data like partial atomic charges that, while easy to extract from DFT simulations, are not physical observables and ubiquitously defined [ 19 ] . To overcome this issue, different methods have been proposed to rely on physical observables, like the electron densities [ 24 ] , the Wannier functions [ 25 ] , and polarization [ 26 , 27 ] , which can be rigorously defined through the modern theory of polarization [ 28 , 29 , 30 ] . These approaches have also been augmented to describe charge relocalization by solving a charge equilibration equation self-consistently, the so-called 4th generation scheme [ 31 , 32 ] . While promising, these procedures require new first-principles calculations to store the extra information needed for the training, hampering the reuse of energy/forces data generated, for example, from previous first-principles calculations of molecular dynamics (FPMD). A different strategy introduces long-range descriptors of atomic coordinates [ 33 ] , also achieved by message-passing neural architectures architectures [ 34 ] . While very promising, especially in their ability to account for macroscopic rearrangement of the electron charge in the system, these models need a training set with sufficiently large simulation cells to sample long-range interactions [ 33 , 27 , 34 ] , preventing the employment of small simulation cells in the training data. This work presents instead a comprehensive first-principles approach designed to integrate long-range interactions seamlessly with existing short-range force fields in polar condensed matter systems. The method accurately calculates total energies, forces, and stress tensors, and it is grounded solely on a select set of well-established first-principles properties of the material in its equilibrium state. II Results II.1 Electrostatic energy The coupling between dipoles is the leading term of the long-range electrostatic interactions in neutral systems. Each dipole \bm{\mu}_{i} generates an electric field \bm{E}(\bm{R}_{i}) decaying as the inverse of the cube of the distance, which interacts with other dipoles as \mathcal{E}=-\frac{1}{2}\sum_{i=1}^{N}\bm{E}(\bm{R}_{i})\cdot\bm{\mu}_{i}, (1) where bold symbols represent vectors and matrices; the \frac{1}{2} in Eq. ( 1 ) prevents double counting when summing over all the atoms. The overall dipole moment \bm{\mu} of the crystal can be rigorously defined through the modern theory of polarization [ 28 , 29 , 30 ] . Here, we can define the atomic dipole \bm{\mu}_{i} as the contribution to the overall dipole moment \bm{\mu} given by the displacement of atom i Z_{i\alpha\beta}=\frac{\partial\mu_{\alpha}}{\partial R_{i\beta}},\qquad\mu_{\alpha}(\bm{R})=\mu^{(0)}_{\alpha}+\sum_{i=1}^{N}\mu_{i\alpha}(\bm{R}), (2) \mu_{i\alpha}(\bm{R})=\sum_{\beta}Z_{i\alpha\beta}(R_{i\beta}-\mathcal{R}_{i\beta}), (3) where \mu_{\alpha} and \mu_{i\alpha} are the \alpha Cartesian component of the dipole moment per cell of the total system and the one of the i -th atom, respectively. Z_{i\alpha\beta} is the Born effective charges tensor, \bm{\mathcal{R}_{i}} is the equilibrium position of the i -th atom, and \bm{\mu}^{(0)} is the total dipole moment when atoms are in the equilibrium position, defined up to a quantum of polarization. We identify the atoms with Latin indices and the Cartesian components with Greek ones. To parametrize the electric field generated by the dipoles of Eq. ( 3 ), we define an auxiliary distribution of spherical charges reproducing a dipole \bm{\mu}_{i} on each atom. In particular, each dipole consist of two charges q of opposite signs and at a distance d as \bm{\mu}=q\bm{d}\;. (4) The modulus of the dipole qd leaves an arbitrary choice on the values for q and d , which, however, only affects terms of the multipole expansions beyond the dipole and, thus, disappears from the final expression for the energy and its derivatives (forces and stresses) in the dipole limit (Appendix C ).

Read the complete text