Hierarchical universal open structure dataset for atomistic modeling (MAD-COSMO)

MAD-COSMO is the collection of almost 1.7 billion atomistic structures from publicly available datasets from which MAD-CORE was selected. Each structure is assigned to its nearest MAD-CORE structure, forming Voronoi cells that follow the MAD-CORE order. Within each cell, structures are ordered using farthest point sampling.

The record contains geometries, energies and non-conservative forces predicted by the ecumetric SPACE model, 512-dimensional features, source identifiers, metadata and filtering diagnostics. It also stores the nearest MAD-CORE structure and its distance at each of the 25 levels. Structures rejected by filters and exact duplicated structures identified by structure hashes are kept and flagged.

Structures and features are stored in matching HDF5 files covering intervals of MAD-CORE cells. No DFT reference calculations are included.