Skip to content

A library imlementing a graph neural network with sparse representation from Code for Kazeev, N., Al-Maeeni, A.R., Romanov, I. et al. Sparse representation for machine learning the properties of defects in 2D materials. npj Comput Mater 9, 113 (2023).

License

HSE-LAMBDA/MEGNetSparse

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

19 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

MEGNetSparse

Installation

pip install MEGNetSparse
  1. You must first install the torch and torch_geometric
  2. The notebook provided in the examples will only work with pymatgen==2023.1.30, so you may need to reinstall it.

Summary

The library implements MEGNet model for sparse representations of crystals with point defects. The essence of the method is depicted in the figure below: Sparse representation construction

Usage

The library consists of two main parts: the first is the construction of a sparse representation of a crystal with point defects, the second is the training of a MEGNet model on this representation.

Sparse representation construction

convert_to_sparse_representation(
    structure,
    unit_cell,
    supercell_size,
    skip_eos=True,
    skip_was=False,
    skip_state=False,
    copy_unit_cell_properties=False
)
  • structure : Structure - the structure to convert to sparse representation
  • unit_cell : Structure - unit cell of base material
  • supercell_size : List[int] - list with three integers to copy a cell along three coordinates
  • skip_eos : bool - if True will not add engineered features aka EOS to properties and will speed up computations
  • skip_was: bool - if True will not add the types of the atoms in the pristine material on the defect sites aka was to properties
  • skip_state : bool - if True will not add global state, which consists of a vector with the atomic numbers of the atoms present in the pristine material
  • copy_unit_cell_properties: bool - if True will also copy unit cell properties in case of name collisions structure properties will be overwritten

return : Structure - sparse representation of the structure

Model training

MEGNetTrainer(
    config,
    device,
)
  • config : dict - template config can be found in examples notebook
  • device : str - device in torch format
MEGNetTrainer.prepare_data(
    self,
    train_data,
    train_targets,
    test_data,
    test_targets,
    target_name,
):
  • train_data : List[Structure] - list of structures in sparse or dense representation. The model is agnostic to the representation and won't do the conversion; if you supply dense structures (aka ordinary pymatgen.Structure) the model will be roughly equal to ordinary MEGNet
  • train_targets : List[float32] - list of targets
  • test_data : List[Structure] - list of structures in sparse or dense representation
  • test_targets : List[float32] - list of targets
  • target_name : str - target name
MEGNetTrainer.train_one_epoch(self)

return : mae on train data, mse on train data

MEGNetTrainer.evaluate_on_test(
    self, 
    return_predictions=False
)

return : if return_predictions=True, mae on test data, predictions else only mae on test data

MEGNetTrainer.predict_structures(
    self, 
    structures_list
)
  • structures_list : List[Structure] - list of structures in the same representation (dense/sparse) as was used for training

return : predictions for structures

MEGNetTrainer.save(self, path)
  • path : str - where to store model data
MEGNetTrainer.load(self, path)
  • path : str - where to load model data from

Citation

Please cite the following paper if you use the library:

Kazeev, N., Al-Maeeni, A.R., Romanov, I. et al. Sparse representation for machine learning the properties of defects in 2D materials. npj Comput Mater 9, 113 (2023). https://doi.org/10.1038/s41524-023-01062-z

About

A library imlementing a graph neural network with sparse representation from Code for Kazeev, N., Al-Maeeni, A.R., Romanov, I. et al. Sparse representation for machine learning the properties of defects in 2D materials. npj Comput Mater 9, 113 (2023).

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages