descriptor.out

The descriptor.out file contains the per-structure or per-atom descriptor values for the input train.xyz file.

This file is written only in prediction mode (prediction 1), with output_descriptor set to 1 or 2. It is not written during training. Descriptor output is not available for model nep_dipole or model nep_polarizability. Each prediction run appends to an existing file; remove or rename it before running if only the new results are wanted. Mode 1 writes one row per structure; mode 2 writes one row per atom.

File format

Each prediction run writes \(N\) rows and \(N_{\rm des}\) columns, where \(N\) is the number of structures or atoms in train.xyz and \(N_{\rm des}\) is the dimension for the descriptor vector.

Within each run’s output, the row index is consistent with the global structure or atom index in the train.xyz file.

Each row contains the normalized descriptor components used by the model. Mode 1 averages these components over the atoms in each structure.

For ordinary NEP, radial components come first, followed by three-body angular components. The remaining invariant blocks follow the switch order in l_max: q222, q1111, q112, q123, q233, and q134, omitting disabled blocks. Within each angular block, the radial index increases from 0 to \(n_\mathrm{max}^\mathrm{A}\). For NEP-spin, the magnetic descriptor components follow the structural components.