FHI-vibes: Ab Initio Vibrational Simulations

The vibrational motion of nuclei determines many important properties of materials, including their thermodynamic equilibrium and non-equilibrium properties. Accurately assessing the nuclear dynamics and the associated material properties is therefore an important task for computational materials scientists in a broad range of sub-fields. Of particular importance are simulation techniques that build on first-principles electronic-structure simulations and thereby allow to systematically investigate the virtually infinite space of materials, including those systems for which little or no experimental data is hitherto available (Curtarolo et al., 2013). This allows one to design novel and improved materials with optimal properties for many applications, e.g., high-performance thermal insulators for gas and airplane turbines (Evans, Clarke, & Levi, 2008), organic semicondcutors with long-term phase stabilities (Salzillo et al., 2016), thermoelectric generators (Snyder & Toberer, 2008), and improved thermal management systems (Tian & Ren, 2019).


Introduction
The vibrational motion of nuclei determines many important properties of materials, including their thermodynamic equilibrium and non-equilibrium properties. Accurately assessing the nuclear dynamics and the associated material properties is therefore an important task for computational materials scientists in a broad range of sub-fields. Of particular importance are simulation techniques that build on first-principles electronic-structure simulations and thereby allow to systematically investigate the virtually infinite space of materials, including those systems for which little or no experimental data is hitherto available (Curtarolo et al., 2013). This allows one to design novel and improved materials with optimal properties for many applications, e.g., high-performance thermal insulators for gas and airplane turbines (Evans, Clarke, & Levi, 2008), organic semicondcutors with long-term phase stabilities (Salzillo et al., 2016), thermoelectric generators (Snyder & Toberer, 2008), and improved thermal management systems (Tian & Ren, 2019).
Essentially, there are two distinct routes towards assessing vibrational properties: In perturbative lattice dynamics techniques, the potential-energy surface on which the nuclei move is approximated with a Taylor expansion around the equilibrium structure. Typically, one starts from a second-order expansion, i.e., the harmonic approximation, which allows for an analytic solution of the equations of motion (Dove, 1993) and thus for a straightforward evaluation of observables (thermodynamic expectation values). Higher-order terms in the Taylor expansion can be accounted for perturbatively. Conversely, molecular dynamics (MD) based approaches account for the full, non-perturbative potential-energy surface without approximating the actual interactions. This requires one to solve the equations of motion numerically by propagating the atoms in time; physical properties can then be extracted as time averages of properly chosen observables (Tuckerman, 2010). Although both lattice dynamics and molecular dynamics techniques aim at computing the same physical observables, the involved methodologies, formalisms, and challenges are quite different. Accordingly, both methodologies also have different strengths and weaknesses: For instance, performing and analyzing MD simulations is typically computationally and conceptually more challenging, whereas perturbative lattice dynamics calculations inherently rely on approximations that are hard to validate.
To date, a variety of different software packages exists at different degrees of sophistication in both fields. Prominent examples are the phonopy code  for performing lattice dynamics calculations using Parlinski's finite-difference formalism (Parlinski, Li, & Kawazoe, 1997) and the i-PI code (Kapil et al., 2019) for performing classical MD and quantum-mechanical path-integral MD simulations. Both packages interface with a variety of first-principles codes like VASP (Kresse & Furthmüller, 1996), QuantumEspresso (Giannozzi et al., 2009), Abinit (Gonze et al., 2020), FHI-aims (Blum et al., 2009), and several others.

Statement of need
To date, there is no software solution that allows for the seamless bridging and interlinking of lattice dynamics and molecular dynamics based approaches, despite the fact that actual material science studies can profit in accuracy and efficiency by exploiting both approaches. For instance, this allows to accelerate MD calculations by starting from harmonic equilibrium configurations (West & Estreicher, 2006), to analyze MD simulations in terms of harmonic phonons (Turney, Landry, McGaughey, & Amon, 2009), to investigate the range of validity of the perturbative expansion used in lattice dynamics (Knoop, Purcell, Scheffler, & Carbogno, 2020), and to efficiently overcome finite size-and time-effects in ab initio Green Kubo simulations of the thermal conductivity (Carbogno, Ramprasad, & Scheffler, 2017). Given the wide opportunities for application, the aspect of integration, i.e., the ability to utilize different methodologies from distinct codes in a flexible fashion using a consistent user interface, is paramount. In particular, this is a prerequisite for automatizing these workflows to enable hierarchical high-throughput screening of whole material classes in a systematic fashion. For example, such a workflow would start from geometry optimizations followed by a study of harmonic properties for many materials, so to single out candidate materials for more involved, fully anharmonic aiMD simulation techniques. Along these lines, let us mention that providing descriptive input and output files is a prerequisite for sharing raw data and results in a transparent and interpretable way in the spirit of open science and the FAIR Principles (Draxl & Scheffler, 2018). On top of that, tracking the provenance (Huber et al., 2020) across different codes facilitates the repurposing and analysis of the obtained data.

Summary
FHI-vibes is a python package that allows for such an integrated workflow. It uses the Atomistic Simulation Environment (ASE) (Hjorth Larsen et al., 2017) as a backend in order to represent materials and to connect to various first-principles codes, to implement methods like geometry optimization and MD, and to connect to external codes like spglib (Togo & Tanaka, 2018), phonopy , phono3py , and hiphive (Eriksson, Fransson, & Erhart, 2019) that implement lattice dynamics techniques. For all these tasks, FHI-vibes provides defined input files and a command line interface to set up and run calculations on local machines and clusters using the slurm submission system. The output is organized in self-contained and descriptive output files that enable a straightforward exchange of the data obtained with different methodologies. For advanced analysis, it provides an API fully compatible with ASE as well as numpy (van der Walt, Colbert, & Varoquaux, 2011), pandas (McKinney, 2011, and xarray (Hoyer & Hamman, 2017); several user-friendly utilities allow to perform the most common postprocessing tasks within the command-line interface, such as providing comprehensive summaries of MD simulations or phonon calculations. (Jain et al., 2015), a workflow management system for running simulation workflows on extensive sets of materials in high-throughput fashion. FHI-vibes is tightly integrated with FHI-aims (Blum et al., 2009) to perform energy and force calculations, but extending the functionality to any calculator available via ASE is straightforward.

FHI-vibes provides a connection to FireWorks
An extensive user guide including a comprehensive list of the available features, tutorials, and a reference documentation is available at vibes.fhi-berlin.mpg.de.