agat 9.1.1

Atomic Graph ATtention networks for predicting atomic energies and forces.

AGAT (Atomic Graph ATtention networks)

Using AGAT

The documentation of AGAT API is available.

Installation

Install with conda environment

• Download the agat_linux_gpu_cu124.yml file.

• Run

  conda env create -f agat_linux_gpu_cu124.yml
  

• Install CUDA and CUDNN [Optional].

  • For HPC, you may load CUDA by checking module av, or you can contact your administrator for help.
  • CUDA Toolkit
  • cuDNN

• More installation options: Customized installation

Quick start

Prepare VASP calculations

Run VASP calculations at this step.

Collect paths of VASP calculations

• We provided examples of VASP outputs at VASP_calculations_example.

• Find all directories containing OUTCAR file:

  find . -name OUTCAR > paths.log
  

• Remove the string 'OUTCAR' in paths.log.

  sed -i 's/OUTCAR$//g' paths.log
  

• Specify the absolute paths in paths.log.

  sed -i "s#^.#${PWD}#g" paths.log
  

Build database

from agat.data import BuildDatabase
if __name__ == '__main__':
    database = BuildDatabase(mode_of_NN='ase_dist', num_of_cores=16)
    dataset = database.build()

Train AGAT model

from agat.model import Fit
f = Fit()
f.fit()

Application (geometry optimization)

from ase.optimize import BFGS
from ase.io import read
from agat.app import AgatCalculator

model_save_dir = 'agat_model'
graph_build_scheme_dir = 'dataset'

atoms = read('POSCAR')
calculator=AgatCalculator(model_save_dir,
                          graph_build_scheme_dir)
atoms = Atoms(atoms, calculator=calculator)
dyn = BFGS(atoms, trajectory='test.traj')
dyn.run(fmax=0.05)

Application (high-throughput prediction)

from agat.app.cata import HtAds

model_save_dir = 'agat_model'
graph_build_scheme_dir = 'dataset'
formula='NiCoFePdPt'

ha = HtAds(model_save_dir=model_save_dir, graph_build_scheme_dir=graph_build_scheme_dir)
ha.run(formula=formula)

Tips:

See API doc for more details. For example:

• Manipulating agat.dataset:
• AGAT molecular dynamics simulations:
• More options for controlling the AGAT training process: docs/sphinx/source/Default parameters.md.

Some default parameters

agat/default_parameters.py; Explanations: docs/sphinx/source/Default parameters.md.

Package structure

Change log

Please check Change_log.md