Data Type for Generation

Before reading this part, make sure the you have already known Structure from Sample Data and Background.

Definition

We divided the features into the following categories:

  1. atom/element feature *

    The properties of atoms/elements themselves.

  2. bond features

    The properties of inter-atomic bonds.

  3. state (overall compound) features *

    The overall properties of the compound, include the properties that embody the overall crystal structure.

  4. crystal structure features (graph features)

    The total of atom/element feature, bond features, state (overall compound) features.

Note

Different from 1,3, The 2,4 is cannot be used directly for sklearn, but suit for torch .

Access

All the Generation tools with convert method for single case, and fit_transform methods for case list.

1. Atom/Element Features

Example Graph 1:

../_images/input_data_type.png

Example Graph 2:

../_images/data_type.jpg

Atom/Element features can be obtained by fetching periodic table data, using you input data. There are two data should offer. (at least one).

  • Your input data. The type could be atom number ( list ) or element name ( dict ) or pymatgen Structure.

    (we have built-in conversion functions of Structure, to directly get all the atomic information in compound).

  • The element periodic table data (optional).

    We have built-in some element periodic table (“ele_table.csv”, “ie.json”, “oe.csv”), To customize your element periodic table. you can offer ( .json , .csv ) file or any python data ( dict , pandas.DataFrame , numpy.ndarray ) in code.

    1. ( .json , dict ) by AtomJsonMap ,

    2. (.csv, pandas.DataFrame , numpy.ndarray ) by AtomTableMap .

    3. And one specialized AtomPymatgenPropMap for fetch data from “pymatgen.core.periodic_table.json”.

Example:

Input atom list

>>> from featurebox.featurizers.atom.mapper import AtomTableMap
>>> tmps = AtomTableMap(search_tp="number")
>>> single_sample = [1,1,1,76,76]
>>> multi_sample = [[1,1,1,76,76],[3,3,4,4]]
>>> a = tmps.convert(single_sample)
>>> b = tmps.transform(multi_sample)

Input element dict

>>> from featurebox.featurizers.atom.mapper import AtomJsonMap
>>> tmps = AtomJsonMap(search_tp="name",return_type="np")
>>> single_sample = [{"H": 2}, {"Po": 1}]
>>> single_sample2 = {"H": 2, "Po": 1}
>>> multi_sample = [[{"H": 2}, {"Po": 1}],  [{"He": 3}, {"P": 4}]] # or
>>> multi_sample2 = [{"H": 2, "Po": 1},  {"He": 3, "P": 4}]
>>> a = tmps.convert(single_sample)
>>> a = tmps.convert(single_sample2)
>>> b = tmps.transform(multi_sample)
>>> b = tmps.transform(multi_sample2)

Input structure type

>>> from featurebox.featurizers.atom.mapper import AtomJsonMap
>>> tmps = AtomJsonMap(search_tp="name",return_type="np")
>>> a = tmps.convert(structurei)
>>> b = tmps.transform(structure_list)
More:

Json Data

2. Bond Features

1. For bond features, use the structure data to extract information. The common structure data include the Structure of Pymatgen, the Atoms of ase, etc. The Structure and Atoms could mutual transform by pymatgen.io.ase.AseAtomsAdaptor.

3. State (overall compound) Features

There are two method to get state (overall compound) features.

  • 1. Information extraction from structure data.

Example:

>>> from pymatgen.core.structure import Structure
>>> structurei = Structure.from_file(r"your_path/featurebox/data/temp_test_structure/W2C.cif")

>>> from featurebox.featurizers.state.state_mapper import StructurePymatgenPropMap
>>> tmps = StructurePymatgenPropMap(prop_name = ["density", "volume", "ntypesp"])
>>> a = tmps.convert(structurei)
>>> b = tmps.transform([structurei]*10)

where the prop_name is the name of properties of in pymatgen , the name of properties is not apply for all compounds, and the data could not a single number.:

prop_name = ["atomic_radius","atomic_mass","number","max_oxidation_state","min_oxidation_state",
"row","group","atomic_radius_calculated","mendeleev_no","critical_temperature","density_of_solid",
"average_ionic_radius","average_cationic_radius","average_anionic_radius",]

  • 2. Combination or mathematical processing of atomic features according to composition ratio.

This is one key method to get state features!!! We can get the results directly or in two step as needed.

  • Get State features directly.

>>> from pymatgen.core.structure import Structure
>>> from featurebox.featurizers.state.statistics import WeightedAverage
>>> structurei =Structure.from_file(r"your_path/featurebox/data/W2C.cif")

>>> from featurebox.featurizers.atom.mapper import AtomTableMap
>>> data_map = AtomTableMap(search_tp="name", n_jobs=1)
>>> wa = WeightedAverage(data_map, n_jobs=1,return_type="df")
>>> x3 = [{"H": 2, "Pd": 1},{"He":1,"Al":4}]
>>> wa.fit_transform(x3)
>>> x4 = [structurei]*5
>>> wa.fit_transform(x4)

More combination operation WeightedSum , GeometricMean , HarmonicMean , WeightedVariance and so on can be found in featurebox.featurizers.state.statistics. More: Pymatgen Data.

  • Get State features by step (Just for compositions with same number of atomic types).

Get the depart element feature first.

>>> from featurebox.featurizers.atom.mapper import AtomJsonMap
>>> from featurebox.featurizers.state.union import UnionFeature
>>> from featurebox.featurizers.state.statistics import DepartElementFeature
>>> data_map = AtomJsonMap(search_tp="name",embedding_dict="ele_megnet.json", n_jobs=1) # keep this n_jobs=1 and return_type="np"
>>> wa = DepartElementFeature(data_map,n_composition=2, n_jobs=1, return_type="pd")
>>> comp = [{"H": 2, "Pd": 1},{"He":1, "Al":4}]
>>> wa.set_feature_labels(["fea_{}".format(_) for _ in range(16)]) # 16 this the feature number of built-in "ele_megnet.json"
>>> couple_data = wa.fit_transform(comp)

Union the depart element feature.

>>> # couple_data is the pd.Dataframe table.
>>> # comp is the atomic ratio of composition.
>>> uf = UnionFeature(comp,couple_data,couple=2,stats=("mean","maximum"))
>>> state_data = uf.fit_transform()

Note

The UnionFeature also could be used for your own table data!

Addition:

There one state features transformer to get Polynomial extension for table.

>>> import numpy as np
>>> from featurebox.featurizers.state.union import PolyFeature
>>> state_features = np.array([[0,1,2,3,4,5],[0.422068,0.360958,0.201433,-0.459164,-0.064783,-0.250939]]).T
>>> state_features = pd.DataFrame(state_features,columns=["f1","f2"],index= ["x0","x1","x2","x3","x4","x5"])
>>> pf = PolyFeature(degree=[1,2])
>>> pf.fit_transform(state_features)
More:

Polynomial Combination, Combination.