features

Construction of features for regression

After obtaining a joined ASE database use construct_feats.py to prepare all slabs and corresponing adsorbates as features for regression of adsorption energies. Currently, two types of features are available:

Zone features applies a template of counting the number of elements in each zone of equivalent atomic positions as described in previous publications. Per default, this includes the nearest neighbors of the binding site and select next-nearest neighbors deemed important to the adsorption energy. This feature scheme requires a separate regression models since the feature vector will vary in length for different sites. These features are saved as a numpy-array with each row representing a feature with the corresponding adsorption energies in the last column.

Graph features converts the structures to graphs consisting of a list of node features and a sparse edge matrix and per default all next-nearest neightbors are included. The element of each node/atom is one-hot encoded and the atomic layer is noted (0 for adsorbates, 1 for surface-layer, 2 for subsurface-layer etc.). Addtionally, a feature called aoi tracks important atomic positions with favourable long-ranged interactions. All edge pair are These features are saved as a list of PyTorch Geometric Data-objects from which following information can be accessed: 'x': Node features 'y': Adsorbtion energy 'edge_index': Sparse edge matrix 'site': Adsorbtion site 'ads': Adsorbate 'ens': Element composition of binding site

Reference energies are loaded from individual databases and all adsorbtion energies are normalized relative to adsorption on pure Pt(111).

ref_dict = {'ontop_OH':ase.db.connect('../data/3x3x5_pt111_ontop_OH.db').get().energy,
			'fcc_O':ase.db.connect('../data/3x3x5_pt111_fcc_O.db').get().energy,
			'slab':ase.db.connect('../data/3x3x5_pt111_slab.db').get().energy
			 }

List which sites/adsorbate combinations should be included. A list of elements are also needed.

project_name = 'agirpdptru'

site_list = ['ontop','fcc']
ads_list = ['OH','O']

surface_elements = ['Ag','Ir','Pd','Pt','Ru']
adsorbate_elements = ['O','H']

Zone feature construction will result in a set of features for each sites/adsorbate combination. Per default, any samples with forces larger than 0.1 eV/Å is not included.

for i in range(len(site_list)):

	## load joined ASE datebase
	db = ase.db.connect(f'../data/{project_name}.db')

	## filename used for pickling
	filename = f'{project_name}_{site_list[i]}_{ads_list[i]}'

	## Construct zoned features and pickle
	print(f'Performing zonefeat construction of {project_name}_{site_list[i]}_{ads_list[i]}')
	samples = db_to_zonedfeats(surface_elements, site_list[i], ads_list[i], 0.1, db, ref_dict)

	with open(filename + '.zonefeats', 'wb') as output:
		pickle.dump(samples, output)

Graph feature construction will make a joined set of features with all available samples.

## Construct graphs and pickle
print(f'Performing graph construction of {project_name}')
samples = db_to_graphs(surface_elements, adsorbate_elements, 2, 0.1, db, ref_dict)
with open(f'{project_name}.graphs', 'wb') as output:
	pickle.dump(samples, output)