Background

Force constants

Commonly descriptions of the vibrational dynamics of a crystal start with a Taylor expansion of the potential energy of the system in terms of the ionic displacements \(u_i\) away from a set of static equilibrium positions \(R_i\)

\[H = H_0 + \Phi_i^\alpha u^\alpha_i + \frac{1}{2}\Phi_{ij}^{\alpha\beta} u^\alpha_i u^\beta_j + \frac{1}{6}\Phi_{ijk}^{\alpha\beta\gamma} u^\alpha_i u^\beta_j u^\gamma_k + \dots\]

Here, the Einstein summation convention applies for all repeated indices. Roman letters denote atom labels while Greek letters denote the Cartesian directions \(\alpha={x,y,z}\). The first term \(H_0\) is a constant energy shift and can thus be chosen as the reference for a particular equilibrium lattice. The second term vanishes since the expansion is carried around the static equilibrium positions, for which the forces vanish by definition.

The expansion coefficients \(\Phi\) are the so-called force constants. They relate the displacements \(u_i\) not only to the potential energy \(H\) of the system but also the atomic forces \(F_i^\alpha\) according to

\[F_i^\alpha = - \Phi_{ij}^{\alpha\beta} u^\beta_j - \frac{1}{2}\Phi_{ijk}^{\alpha\beta\gamma} u^\beta_j u^\gamma_k - \dots\]

The force constants encode essential information about the vibrational properties of a system and form the basis for a systematic treatment of the lattice dynamics in terms of phonons. An extensive description of phonon theory can be found for example in the classic text books by Ziman [Zim60], Wallace [Wal98], or Born and Huang [BorHua98]. A more recent review including applications has been provided by Fultz [Ful10].

Calculating force constants

The calculation of the force constants commonly represents the computationally most demanding task in a theoretical analysis of the lattice dynamics of a system. The force constants can in principled be obtained via a direct (real space) approach using simple finite differences. For example in the case of the second order force constant matrix (using for simplicity a forward difference)

\[\Phi_{ij}^{\alpha\beta} = \frac{\partial^2 H}{\partial u_i^\alpha \partial u_j^\beta} = - \frac{\partial F_i^\alpha}{\partial u_j^\beta} \approx - \frac{\partial F_i^\alpha(h) - F_i^\alpha(0)}{h},\]

where \(h\) denotes a finite displacement of atom \(j\) along the Cartesian direction \(\beta\). While this approach can be readily extended to higher order differences, the number of calculations needed quickly increases even if the crystal symmetry and intrinsic symmetries of the force constant tensors are taken into account.

_images/clusters.svg

Number of orbits, clusters, and parameters as a function of cutoff radius for FCC Al.

To circumvent this problem hiPhive takes a somewhat different approach. Instead of systematically computing all terms (equivalent to an enumeration of the terms) it treats the elements of the force constant tensors as fit parameters and uses machine learning algorithms to obtain sparse solutions. Specifically, the task of finding the fit parameters \(\vec{a}\) is cast in the form of a linear equation \(\vec{f}=\matrix{M}\vec{a}\), where \(\vec{f}\) is a vector of target forces and \(\matrix{M}\) is the fit matrix.

This is in the spirit of the compressive sensing approach [CanWak08] previously advocated for lattice dynamics [AndSanAsp12], [ZhoNieXia14], [SanAndAsp15]. The present approach is, however, not restricted to a specific optimization strategy. Rather hiPhive is interfaced with scitkit-learn and provides a simple interface to several powerful optimization algorithms, including but not limited to LASSO, Bayesian ridge regression, Bayesian Automatic Relevance Determination regression and conventional \(\ell_2\)-norm based local minimization techniques. Furthermore, it is straightforward to include other optimization algorithms available via Python libraries such as e.g., scipy or TensorFlow.

hiPhive exploits crystal symmetries (via Spglib) and sum rules to minimize the number of independent parameters and has been designed for flexibility while maintaining a high execution speed (aided by NumPy and Numba). The implementation focuses on the handling of the force constants and provides interfaces to other codes and libraries for carrying out further analysis, including

  • ASE (e.g., for molecular dynamics simulations)
  • dynasor (e.g., for obtaining temperature dependent phonon frequencies and life times)
  • phonopy (for analyzing harmonic phonons and properties derived thereof)
  • phono3py (for phonon lifetimes in the perturbative limit and Boltzmann transport theory)
  • shengBTE (for phonon lifetimes in the perturbative limit and Boltzmann transport theory)