Scikit-learn integration¶

This page gives an overview of classes and methods that can be used with the scikit-learn python package.

List of structure transformers¶

The structure transformer classes are based on scikit-learn’s BaseEstimator class and can be combined with other transformer or estimator classes by defining pipelines (more details on scikit-learn’s pipeline framework can be found here).

Python class	Description
`StructureCompositionTransformer`	Transforms the structures into feature vectors consisting of the elemental concentrations.
`StructureDensityTransformer`	Transforms the perdiodic structures into feature vectors of the atomic density (nr atoms per cell volume) of each element.
`StructureCoordinationTransformer`	Transforms the structures into feature vectors consisting of the coordination number and distance to neighbouring atoms for each element pair.
`StructureChemOrderTransformer`	Transforms the structures into their Warren Cowley like order parameters introduced in doi:10.1103/PhysRevB.96.024104.
`StructureFFPrintTransformer`	Transforms the periodic structures into their F-Fingerprint as defined in doi:10.1063/1.3079326.
`StructurePRDFTransformer`	Transforms structures the into their partial radial distribution functions (pRDF) as defined in doi:10.1103/PhysRevB.89.205118.
`StructureMatrixTransformer`	Transforms the structures into different types of interaction matrices as defined in doi:10.1002/qua.24917. The implementation is based on the dscribe python package.
`StructureACSFTransformer`	Transforms the structures into their ACSF descriptor as defined in doi:10.1063/1.3553717. The implementation is based on the dscribe python package.
`StructureSOAPTransformer`	Transforms the structures into their SOAP descriptor as defined in doi:10.1103/PhysRevB.87.184115. The implementation is based on the dscribe python package.
`StructureMBTRTransformer`	Transforms the structures into their many-body tensor representation (MBTR) as defined in doi:10.1088/2632-2153/aca005. The implementation is based on the dscribe python package.

List of custom metrics and kernels¶

The following metrics and kernels are built on-top of the StructureFFPrintTransformer and can be used for different ML models:

The ffprint_cosine function allows the metric based on the cosine distance originally introduced in doi:10.1063/1.3079326 for the F-Fingerprint descriptor to be used with e.g. the k-neighbours model.
The krr_ffprint_cosine function calculates a kernel directly derived from the F-Fingerprint cosine distance to be used with the kernel ridge regression model.
The krr_ffprint_laplace kernel implements the laplacian kernel of the cosine distance for the F-Fingerprint descriptor to be used with the kernel ridge regression model.