Overview

κALDo is a versatile and scalable open-source software to compute phonon transport in crystalline and amorphous solids. It features real space QHGK calculations and three different solvers of the linearized BTE: direct inversion, self-consistent cycle, and RTA.

The algorithms are implemented using linear algebra operations on tensors, to take advantage of multithreading on GPU and CPU using Numpy, Tensorflow and optimized tensor libraries.

Using the Atomic Simulation Environment, κALDo can calculate the IFCs using several ab initio and molecular dynamics codes, thus enabling the use of first-principles DFT, empirical forcefields and semi-empirical tight- binding. A native LAMMPS interface is also available in the USER-PHONON package. Finally, through seamless integration with the hiPhive package, the IFC calculation can take advantage of compressing-sensing machine learning algorithms

The code is released open-source for the community to use and contribute with edits and suggestions. It is designed on modern software best practices, and we hope to provide a development platform to implement new theory and methods.

Features

Below we illustrate the main features of the code.

Forcefields. Using the Atomic Simulation Environment, kALDo can calculate the interatomic force constants using several ab-initio and molecular dynamics codes. A native LAMMPS interface is also available in the USER-PHONON package. Finally, through a seamless integration with the Hiphive package, the IFC calculation can take advantage of compressingsensing machine learning algorithms.

CPUs and GPUs. Multithread implementation on CPUs and GPUs. The algorithms are implemented using linear algebra operations on tensors, to take advantage of multithreading on GPU and CPU using Numpy, Tensorflow and optimized tensor libraries.

Scalable. In a system of N atoms and N_k k points. kALDo uses (3N)^2 floating point numbers to save the state of the system when using QHGK, (Nk x 3N)^2 for the full solution of the BTE and Nk x 3N^2 when using BTE-RTA.

Performance. The bottleneck, and slow part of ALD simulations is the calculation of the phonons lifetimes and the scattering matrix. This step requires projecting the interatomic potential on 3N phonons modes and the algorithm scales like (Nk x 3N)^3, because of the 3 projections on phonons modes. In kALDo such algorithm is implemented as 2 Nk x 3N tensor multiplications of size Nk x (3N)^2 for BTE calculations while 3N^2 for QHGK.

Open-Source. Free for the community to use and contribute with edits and suggestion. It is designed on modern software best practices, and we hope to provide a development platform to implement new theory and methods.

fa1e0fd9d4464dac80406bbe24756a30