FCC NiFeCrAlTiSi Cluster Expansion Training Dataset

These three datasets of atomic structures in the NiFeCrAlTiSi composition space were created as part of our development of our novel instability-avoiding active learning workflow for training cluster expansions. Each dataset is provided in a plain text .json file to ensure future compatibility and in a more performant, binary ASE[1] .db format. The first dataset, ‘sis’, contains the complete set of 47,720 symmetrically inequivalent FCC lattice decorations which have 6 or fewer atoms in the 6-element composition space. The symmetrically inequivalent structures (SISs) have been uniquely tagged by the number of atoms in each structure (i.e. SIS size) and an index number indicated nth structure of that SIS size. Due to a propagated typo, these tags are listed as ‘orbit_size’ and ‘orbit_index’ in all datasets. The second dataset, ‘relaxed_structures.full’, contains the final DFT-relaxed structures after a subset of their corresponding SISs were relaxed till they met the structure optimization criteria. The relaxed structures have attached caches of their DFT-level total energies, atom forces, atom-projected, and total magnetic moments. Additional tags provide the energy per atom above the convex hull (i.e. hull energy) and the mixing enthalpies. The third an last dataset, ‘mapped_structures.full’, contains the relaxed structures where we have been able to map them to the nearest ideal FCC lattice. The details of the mapping process that are stored as tags are the mean atomic displacement as ‘dravg’, the maximum atomic displacement as ‘drmax’, and the volumetric strain as ‘volumetric_strain’. The hull energy and mixing enthalpy tags are carried over from the relaxed structures to simplify model training. All datasets contain Boolean selection number tags from ‘selection_0’ to ‘selection_10’ that indicate if a SIS was selected (true) or excluded (false) during the corresponding active learning loop number.