Default parameters

The default parameters that control the database construction, model training, high-throughput prediction …

You can import such parameters from agat module.

from agat.default_paramters import default_elements, default_build_properties, default_data_config, default_train_config, default_high_throughput_config

Or you can read the source code.

default_elements

Elements used to build graph. A list of elements that are used to encode atomic features.

['Ac', 'Ag', 'Al', 'Am', 'Ar', 'As', 'At', 'Au', 'B',  'Ba',
'Be', 'Bh', 'Bi', 'Bk', 'Br', 'C',  'Ca', 'Cd', 'Ce', 'Cf',
'Cl', 'Cm', 'Cn', 'Co', 'Cr', 'Cs', 'Cu', 'Db', 'Ds', 'Dy',
'Er', 'Es', 'Eu', 'F',  'Fe', 'Fl', 'Fm', 'Fr', 'Ga', 'Gd',
'Ge', 'H',  'He', 'Hf', 'Hg', 'Ho', 'Hs', 'I',  'In', 'Ir',
'K',  'Kr', 'La', 'Li', 'Lr', 'Lu', 'Lv', 'Mc', 'Md', 'Mg',
'Mn', 'Mo', 'Mt', 'N',  'Na', 'Nb', 'Nd', 'Ne', 'Nh', 'Ni',
'No', 'Np', 'O',  'Og', 'Os', 'P',  'Pa', 'Pb', 'Pd', 'Pm',
'Po', 'Pr', 'Pt', 'Pu', 'Ra', 'Rb', 'Re', 'Rf', 'Rg', 'Rh',
'Rn', 'Ru', 'S',  'Sb', 'Sc', 'Se', 'Sg', 'Si', 'Sm', 'Sn',
'Sr', 'Ta', 'Tb', 'Tc', 'Te', 'Th', 'Ti', 'Tl', 'Tm', 'Ts',
'U',  'V',  'W',  'Xe', 'Y',  'Yb', 'Zn', 'Zr']

default_build_properties

A dictionary defines which properties will be built into the graph.

Parameter	Default value	Alternative(s)	Explanation
energy	True	False	Include total energy when building graphs.
forces	True	False	Include atomic forces when building graphs.
cell	True	False	Include structural cell when building graphs.
cart_coords	True	False	Include Cartesian coordinates when building graphs.
frac_coords	True	False	Include Fractional coordinates when building graphs.
constraints	True	False	Include constraint information when building graphs.
stress	True	False	Include Virial stress when building graphs.
distance	True	False	Include distance between connected atoms when building graphs.
direction	True	False	Include unit vector between connected atoms when building graphs.
path	False	True	Include file path of each graph corresponding to DFT calculations when building graphs.

default_data_config

A dictionary defines how to build a database.

Parameter	Default value	Alternative(s)	Explanation
species	default_elements above	A list of element symbols	A list of elements that are used to encode atomic features.
path_file	‘paths.log’	str	A file of absolute paths where OUTCAR and XDATCAR files exist.
build_properties	default_build_properties above	See [default_build_properties](#default_build_properties )	Properties needed to be built into graph.
topology_only	False	True	Build graph with topology connections only. The energy, forces, cell, and stress will not be included. This setting has higher priority than [default_build_properties](#default_build_properties )
dataset_path	‘dataset’	A str	A directory contains the database.
mode_of_NN	‘ase_natural_cutoffs’	‘ase_natural_cutoffs’, ‘pymatgen_dist’, ‘ase_dist’, and ‘voronoi’	The mode of how to detect connection between atoms. Note that pymatgen is much faster than ase.
cutoff	5.0	A float	Cutoff distance to identify connections between atoms. Deprecated if mode_of_NN is 'ase_natural_cutoffs'
load_from_binary	False	True	Read graphs from binary graphs that are constructed before. If this variable is True, these above variables will be depressed.
num_of_cores	2	int	How many cores are used to extract vasp files and build graphs.
super_cell	False	True	When building graphs, small cell may have problems to find neighbors. Specify this parameter as True to repeat cell to avoid such problems
has_adsorbate	False	True	Include adsorbate information when building graphs. For now, only H and O atoms are considered as adsorbate atoms.
keep_readable_structural_files	False	True	Massive number of structural files (POSCARs) under dataset_path are generated when building graphs, you can choose to keep them or not.
mask_similar_frames	False	True	In VASP calculations, the energy optimization generate many frames that have similar geometry and total energies, you can extract only some of them by specifying this parameter and energy_stride below.
mask_reversed_magnetic_moments	False	float	Frames with atomic magnetic moments lower than this value will be masked.
scale_prop	False	True	Scale the properties. This function seems to be deprecated. I need to double-check the source code first, so do not use it.

default_train_config

A dict determines how to train the AGAT model.

Parameter	Default value	Alternative(s)	Explanation
verbose	1	0, 1	Output verbosity. 0: test output; 1: Validation and test output; 2: train, validation, and test output.
dataset_path	‘dataset’	A str	A directory contains the database.
model_save_dir	‘agat_model’	directory name, str	A directory to save the well-trained model.
epochs	1000	int	Number of training epochs.
output_files	‘out_file’	str	A directory to store ouputs of true and predicted properties.
device	‘cuda:0’	‘cpu’	Device to train the model. Use GPU cards to accerelate training.
validation_size	0.15	float, 0<validation_size<1	Determines the proportion of the dataset to be included in the validation split.
test_size	0.15	float, 0<validation_size<1	Determines the proportion of the dataset to be included in the test split.
early_stop	True	False	Implement early stop or not. If this is True, the training will be terminated after a specified number of epochs without model improvement. If this is False, the model weights will be saved every epoch.
stop_patience	300	int	Activated when early_stop=True. The training will be terminated after stop_patience epochs without model improvement.
head_list	[‘mul’, ‘div’, ‘free’]	list	A list of attention mechanisms. See agat/model/model.py.
gat_node_dim_list	[len(default_elements), 100, 100, 100]	list	Node dimensions of AGAT Layer.
energy_readout_node_list	[len(head_list)*gat_node_dim_list[-1], 100, 50, 30, 10, 3, 1]	list	A list of node dimensions of energy readout layers.
force_readout_node_list	[len(head_list)*gat_node_dim_list[-1], 100, 50, 30, 10, 3, 3]	list	A list of node dimensions of force readout layers.
stress_readout_node_list	[len(head_list)*gat_node_dim_list[-1], 100, 50, 30, 10, 3, 6]	list	A list of node dimensions of stress readout layers.
bias	True	False	Add bias or not to the neural networks.
negative_slope	0.2	float	This specifies the negative slope of the LeakyReLU activation function.
criterion	nn.MSELoss()	torch.nn loss functions	Creates a criterion that measures the mean squared error (squared L2 norm) between each element in the input x and target y.
a	1.0	float	The importance of energy loss in the total loss function. See agat/model/fit.py.
b	1.0	float	The importance of force loss in the total loss function. See agat/model/fit.py.
c	0.0	float	The importance of stress loss in the total loss function. See agat/model/fit.py.
learning_rate	0.0001	float	The learning rate of torch.optim.Adam optimizer.
weight_decay	0.0	float	The weight decay of torch.optim.Adam optimizer.
batch_size	64	int	Training batch size.
val_batch_size	400	int	Batch size when validation and test.
transfer_learning	False	True	Turn on the transfer learning when True. (Deprecated)
trainable_layers	-4	negative int	tail trainable_layers layers are trainable, other layers are freezed. (Deprecated)
mask_fixed	False	True	Mask fixed atoms or not. When True, the atomic forces of fixed atoms will not be included in the loss function. (Deprecated)
tail_readout_no_act	[3,3,3]	list	The tail tail_readout_no_act layers will have no activation functions. The first, second, and third elements are for energy, force, and stress readout layers, respectively.
adsorbate_coeff	20.0	float	Indentify and specify the importance of adsorbate atoms with respective to surface atoms. zero for equal importance.

default_ase_calculator_config.

See bfgs for more details.

Parameter	Default value	Alternative(s)	Explanation
fmax	0.1	float	Convergence criterion of atomic forces. Details: ase optimizer
steps	200	int	Maximum iteration steps.
maxstep	0.05	float	maximum distance an atom can move per iteration, unit is Å.
restart	None	str	Pickle file used to store hessian matrix.
restart_steps	0	int	Restart optimization if the optimization cannot converge.
perturb_steps	0	int	Number of perturbated steps. AGAT may have issues in converging BFGS, perturbating atomic positions may help the convergence.
perturb_amplitude	0.05	float	Perturbation amplitudes if erturb_steps larger than 1.
out	None	str	Base name of the output log and traj.

default_high_throughput_config

Settings for the high-throughput predictions.

Parameter	Default value	Alternative(s)	Explanation
model_save_dir	agat_model	str, a directory name	A directory for loading the well-trained model from.
opt_config	default_ase_calculator_config	dict	Settings for ase.optimize.BFGS structural optimizer.
calculation_index	0	str	To label the calculation outputs.
fix_all_surface_atom	False	True	Fix all surface atoms or not.
remove_bottom_atoms	False	True	Remove the bottom atoms or not.
save_trajectory	False	True	Keep the optimization trajectory.
partial_fix_adsorbate	False	True	Partially fix the adsorbate freedom.
adsorbates	[‘H’]		keys of agat/lib/adsorbate_poscar.py
sites	[‘ontop’]	list	A list of adsorption sites. See agat/app/cata/generate_adsorption_sites.py
dist_from_surf	1.7	float	Distance between adosrbate and surface. Unit: angstrom.
using_template_bulk_structure	False	True	Using template to build the surface model. If this is True, you need to prepare a POSCAR_temp file in the working directory.
graph_build_scheme_dir	‘dataset’	A directory name. str	A directory storing the graph_build_scheme.json file. This file is generated when building the database, and is normally saved in the default_data_config['dataset_path'].
device	‘cuda’	str	Determines the device for the model prediction (forward).

default_hp_dft_config

Default settings for the high-throughput DFT calculations.

To be continued…