Skip to content

DATASET.md

Latest commit

 

History

History
67 lines (48 loc) · 1.62 KB

DATASET.md

File metadata and controls

67 lines (48 loc) · 1.62 KB

ELFNet Dataset Notes

Raw training arrays are stored as Git LFS archive assets under release/, not as regular Git blobs. See DATA_RELEASES.md for the packaged datasets and checksums. Custom training data should use the same paired full-grid SAD/ELF NumPy convention.

Required files:

<stem>_sad.npy
<stem>_elf.npy

Optional files:

<stem>_sym.npy

The current full-grid ELFPredictor ignores _sym.npy files.

Full-Grid Convention

Each *_sad.npy and *_elf.npy pair must have identical shape for a given structure. The loader yields complete unit-cell grids shaped:

sad: (1, D, H, W)
elf: (1, D, H, W)

There is no patch extraction. Shape-bucketed training groups samples by exact grid shape.

SAD Definition

The SAD input is a project-defined superposed atomic density:

  1. neutral spherical atomic density tables are loaded from packaged .pkl files;
  2. each atom contributes a periodic neutral density centered at its fractional position;
  3. the grid is normalized to the configured total valence electron count;
  4. the density is scaled into the ELFNet training convention.

This SAD is not VASP's internal ICHARG=2 density.

ELF Definition

The target ELF grid is parsed from VASP ELFCAR volumetric output and stored as a float32 NumPy array. Values are expected to be ELF-like and typically lie in the range [0, 1].

Training Example

elfnet-train /path/to/paired_arrays \
  --epochs 100 \
  --batch 32 \
  --batching shape \
  --val-frac 0.05 \
  --lambda-cdf 0.05

The trainer expects all required *_sad.npy and *_elf.npy files to be in the same directory.