This package builds ACE interatomic potentials using Hyperactive Learning (HAL) using the ACE.jl Julia software package.

1. install julia 1.7.x and python 3.8 or 3.9 (with python ase, scikit-learn, matplotlib and numpy installed)
2. run julia command (add Pkg.activate(".") for local project)

using Pkg; pkg"registry add https://github.com/JuliaRegistries/General"; pkg"registry add https://github.com/JuliaMolSim/MolSim.git"; pkg"add ACE1, ASE, JuLIP, LinearAlgebra"

3. install julia python package to set up Python -> Julia connection

python -m pip install julia

python -c "import julia; julia.install()"

4. clone this repo and make sure that HAL_lib can be found in your PYTHONPATH

Example scripts are provided in the tutorials folder. A HAL run requires the following main components:

• A Python ASE calculator to perform QM/DFT calculations
• ACE parameters for linear ACE basis in Julia (correlation order, polynomial degree)
• HAL parameters for sampling strategy (biasing strength, uncertainty tolerance, temperature, pressure)

Threading is advised to be enabled by setting environment JULIA_NUM_THREADS for Julia threading and MKL_NUM_THREADS/OPENMP_NUM_THREADS/BLAS_NUM_THREADS to enable numpy/scikit-learn threading

If using this code please reference

@misc{van2022hyperactive,
  doi = {10.48550/ARXIV.2210.04225},
  url = {https://arxiv.org/abs/2210.04225},
  author = {van der Oord, Cas and Sachs, Matthias and Kov{\'a}cs, D{\'a}vid P{\'e}ter and Ortner, Christoph and Cs{\'a}nyi, G{\'a}bor},
  title = {Hyperactive Learning (HAL) for Data-Driven Interatomic Potentials},
  publisher = {arXiv},
  year = {2022},
}