Introduction to structural properties

This section provides an overview of the concepts underlying the usage of chemiscope. The tool is designed to help navigating structure-property maps, which are 2d- or 3d-dimensional embeddings that reflect how variations in atomic structure influence materials properties.

Chemiscope works with two kinds of entities: full structures and atom-centered environments. A structure contains all atoms in a configuration, e.g. the unit cell of a crystal or the geometry of a molecule. An environment consists of a set of neighbors that surround a central atom within a chosen cutoff radius. In both cases, these entities are fully defined by the positions of the atoms and their chemical elements.

Each structure or environment can be associated with properties. Some of these, such as energy, forces, density, or NMR shielding, represent physical properties that describe the system. Others are structural representations that encode the geometry itself. These include a wide family of descriptors, ranging from traditional hand-crafted functions, such as Behler-Parrinello symmetry functions, to more recent approaches such as SOAP, Atomic Cluster Expansion (ACE), or latent embeddings extracted from graph neural network potentials (e.g. PET-MAD). These representations are usually high-dimensional vectors, hard to visualize and interpret. For this reason, it is common to apply dimensionality reduction algorithm, such as PCA, sketch-map, PCovR, etc. These methods compress the descriptor into a small number of components that can be visualized on a map. The interpretability of the low-dimensional embedding depends both on the representation that was originally used and on the algorithm used to reduce its dimensionality.

Chemiscope provides an interface that links each structure or environment to a point in a reduced-dimensionality map and to a 3d-dimensional visualization of its atomic configuration.

Illustration of the process used to create structural properties from a molecule.

Chemiscope is completely agnostic with respect to how properties and structural representations are generated.

There are two main ways to prepare data for chemiscope. One option is to compute structural descriptors or low-dimensional features using external tools, such as scikit-matter or ASAP, and then load them into chemiscope. The other option is to let chemiscope compute these representations automatically, using chemiscope.explore().