Awesome
ElMD
The Element Movers Distance (ElMD) is a similarity measure for chemical compositions. This distance between two compositions is calculated from the minimal amount of work taken to transform one distribution of elements to another along the modified Pettifor scale.
This repository provides the reference implementations as described in our paper "The Earth Movers Distance as a metric for the space of inorganic compositions".
If you wish to compute this metric between lots of compositions, the ElM2D high-performance library may be more useful and can be found at www.github.com/lrcfmd/ElM2D.
We recommend installation via pip
pip install ElMD
For python 3.8+, due to known library conflicts it is reccomended to install ElMD separate to its dependencies
pip install ElMD --no-deps
pip install numpy # if necessary
pip install numba # Gives significant speedup, but can cause dependency issues with other libraries
Usage
For simple usage initiate an object with its compositional formula
> from ElMD import ElMD
> x = ElMD("CaTiO3")
Calculate the distance to a second object with the elmd method.
> x.elmd("SrTiO3")
0.2
If the assignment plan (how each element in the source composition is mapped to the target composition) is required, this may be returned by setting the return_assignments flag in the elmd method.
> x.elmd("SrTiO3", return_assignments=True)
(0.2, array([0.2, 0. , 0. , 0. , 0.2, 0. , 0. , 0. , 0.6]))
If the mod_petti elemental scale is suitable and no assignment plan is required, a significantly faster EMD algorithm may be used by setting metric="fast"
> x = ElMD("CaTiO3", metric="fast")
> x.elmd("SrTiO3")
0.2
The compositional parser can handle user defined values of x when this is applicable.
latp_02 = ElMD("Li1+xAlxTi2-x(PO4)3", x=0.2) # Li1.2Al0.2Ti1.8(PO4)3
latp_03 = ElMD("Li1+xAlxTi2-x(PO4)3", x=0.3) # Li1.3Al0.3Ti1.7(PO4)3
Alternate chemical scales may be accessed via the "metric" argument, e.g.
> x = ElMD("CaTiO3", metric="atomic")
> x.elmd("SrTiO3")
3.6
The elmd() method is overloaded to take two strings, and may be imported directly. The choice of metric is specified with metric
from ElMD import elmd
> elmd("NaCl", "LiCl")
0.5
> elmd("NaCl", "LiCl", metric="magpie")
0.688539
The EMD function can also be called directly, with the input being two vectors of distributions and the associated distance matrix between them.
from ElMD import EMD
> EMD([0.5, 0.5], [0.5, 0.5], [[1, 90], [89, 0]])
0.5
Elemental Similarity
You may use either traditional discrete scales or machine learnt representations for each element. In this instance a vector has been generated for each element, and the distance between elements (not compositions!) is the Euclidean distance.
Due to the disparity in magnitudes of some of these values, a select few have additionally been scaled.
Linear:
Chemically Derived:
Machine Learnt:
Random Numbers:
The Euclidean distance between these vectors is taken as the measure of elemental similarity.
> x = ElMD("NaCl", metric="magpie")
> x.elmd("LiCl")
46.697806
> x = ElMD("NaCl", metric="magpie_sc")
> x.elmd("LiCl")
0.688539
The feature dictionary can be accessed through the periodic_tab attribute:
> featurizingDict = ElMD(metric="magpie).periodic_tab
> featurizingDict["Na"]
[2.0, 22.98976928, 370.87, 1.0, 3.0, 166.0, 0.93, 1.0, 0.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 0.0, 1.0, 29.2433333333, 0.0, 0.0, 229.0]
Featurizing
Whilst not the initial purpose, a compositional based feature vector may be generated from ElMD objects should you require it. This is a mean pooling of the weighted composition feature matrix.
Note that this vector representation is not used at any point during the ElMD distance calculation and is provided solely for convenience.
We construct this by taking the dot product of the ratios of each element with the features of these elements. Pass the argument feature_pooling="mean" to divide by the total number of elements in the compound.
feature_vector = np.dot(ratio_vector, element_feature_matrix)
This is accessed through the feature_vector attribute.
# For single element compositions, equivalent to x.periodic_tab["Cl"]
> x = ElMD("Cl", metric="magpie")
> x.feature_vector
array([ 94. , 35.453 , 171.6 , 17. , 3. , 102. ,
3.16 , 2. , 5. , 0. , 0. , 7. ,
0. , 1. , 0. , 0. , 1. , 24.4975,
2.493 , 0. , 64. ])
# Aggregate vector by each elements contribution
> x = ElMD("NaCl", metric="magpie").feature_vector
array([ 48. , 29.22138464, 271.235 , 9. ,
3. , 134. , 2.045 , 1.5 ,
2.5 , 0. , 0. , 4. ,
0.5 , 0.5 , 0. , 0. ,
1. , 26.87041667, 1.2465 , 0. ,
146.5 ])
A feature vector of length 8076 can be generated by concatenating the weighted mean, min, max, range, and standard deviation across each available elemental feature across all featurizing dictionaries for each element in the composition by calling the full_feature_vector() method.
> x = ElMD("NaCl").full_feature_vector()
When using 1D unpooled elemental vectors, these may be mapped to the associated chemical formula using the vec_to_formula method:
x = ElMD("CaTiO3")
y = ElMD("NaCl")
print(x.pretty_formula)
print(x.vec_to_formula(x.feature_vector)) # Same as above
print(y.vec_to_formula(x.feature_vector)) # Same as above
Citing
If you would like to cite this code in your work, please use the Chemistry of Materials reference
@article{doi:10.1021/acs.chemmater.0c03381,
author = {Hargreaves, Cameron J. and Dyer, Matthew S. and Gaultois, Michael W. and Kurlin, Vitaliy A. and Rosseinsky, Matthew J.},
title = {The Earth Mover’s Distance as a Metric for the Space of Inorganic Compositions},
journal = {Chemistry of Materials},
volume = {32},
number = {24},
pages = {10610-10620},
year = {2020},
doi = {10.1021/acs.chemmater.0c03381},
URL = {
https://doi.org/10.1021/acs.chemmater.0c03381
},
eprint = {
https://doi.org/10.1021/acs.chemmater.0c03381
}
}
Issues
Please feel free to post any questions or comments as issues on this GitHub page.