Hierarchical universal open structure dataset for atomistic modeling. MAD-CORE

MAD-CORE is an ordered coreset of 2²⁴ structures selected from MAD-COSMO, a collection of 1.7 billion structures from publicly available atomistic datasets. Structures are stored in selection order, so the first 2^k structures form the coreset of level k. 

The record contains geometries, model-predicted energies and forces, the features used for the selection, ids, source metadata and diagnostics, as gzipped extended XYZ and HDF5 files, and a chemiscope map of the first 2²⁰ structures.

It also includes FHI-aims DFT reference calculations at LDA and PBE for about 79,000 structures from four selections (the first 25,000 MAD-CORE structures, 25,000 structures sampled with Voronoi cells and 31,000 structures drawn uniformly from MAD-COSMO) and  21,854 r2SCAN structures used in the validation experiments.