Test data for MALA

This repository contains data to test, develop and debug MALA and MALA based runscripts. If you plan to do machine-learning tests ("Does this network implementation work? Is this new data loading strategy working?"), this is the right data to test with. It is NOT production level data!

Be2

Contains DFT calculation output from a QuantumEspresso calculation for a beryllium cell with 2 atoms, along with input scripts and pseudopotential to replicate this calculation. LDOS files are usually large, therefore this reduced example samples the LDOS somewhat inaccurately, in order to reduce storage size. The energy grid for the LDOS is 11 entries long, starting at -5 eV with a spacing of 2.5 eV. For LDOS and descriptors, 4 snapshots are contained. In detail, the following data files can be found:

File Name	Description
recreate_data/	Input scripts for QE
cubes/	.cube files for the local density of states
Be.pbe-n-rrkjus_psl.1.0.0.UPF	Pseudopotential used for the QE calculation
Be_snapshot0.dens.npy	Electronic density numpy array (snapshot 0)
Be_snapshot.dens.h5	Electronic density (HDF5 format, see details below)
Be_snapshot0.dos.npy	Density of states numpy array (snapshot 0)
Be_snapshot0-3.out	Output file of QE. calculation
Be_snapshot0-3.in.npy	Bispectrum descriptors numpy array
Be_snapshot0-3.out.npy	Local density of states numpy array
Be_snapshot0-3.in.h5	Bispectrum descriptors (HDF5 format)
Be_snapshot0-3.out.h5	Local density of states (HDF5 format)
Be_model.zip	MALA trained model archive for examples and tests

numpy format files

Bispectrum descriptors of length 91 on 18 x 18 x 27 real space grid.

Note

In the last dimension of length 94, the first 3 entries are the grid coordinates / indices (an artifact of the SNAP vector generation). The actual features are snap_array[..., 3:].

>>> np.load('Be2/Be_snapshot1.in.npy').shape
(18, 18, 27, 94)

LDOS (11 points) on 18 x 18 x 27 real space grid.

>>> np.load('Be2/Be_snapshot1.out.npy').shape
(18, 18, 27, 11)

Density of states (only provided for snapshot 0):

>>> np.load('Be2/Be_snapshot0.dos.npy').shape
(11,)

Density for snapshot 0 on a 18 x 18 x 27 real space grid. The extra dimension can be ignored, i.e. use d=np.load(...); d[..., -1] to squeeze the shape to (18, 18, 27).

>>> np.load('Be2/Be_snapshot0.dens.npy').shape
(18, 18, 27, 1)

openPMD-based files

MALA supports the openPMD format, so we also provide data in that format here.

$ h5ls -r Be_snapshot0.in.h5 | grep Dataset | sort -V
/data/0/meshes/Bispectrum/0  Dataset {18, 18, 27}
/data/0/meshes/Bispectrum/1  Dataset {18, 18, 27}
...
/data/0/meshes/Bispectrum/93 Dataset {18, 18, 27}

$ h5ls -r Be_snapshot0.out.h5 | grep Dataset | sort -V
/data/0/meshes/LDOS/0     Dataset {18, 18, 27}
/data/0/meshes/LDOS/1     Dataset {18, 18, 27}
...
/data/0/meshes/LDOS/10    Dataset {18, 18, 27}

For the density, the snapshot number 0 is encoded in the name /data/0.

$ h5ls -r Be_snapshot.dens.h5 | grep Dataset
/data/0/meshes/Density/0 Dataset {18, 18, 27}

To understand the naming scheme, we can use openPMD's introspection tool:

$ openpmd-ls Be_snapshot.dens.h5
openPMD series: Be_snapshot.dens
openPMD standard: 1.1.0
openPMD extensions: 0

data author: ...
data created: 2023-05-23 15:37:18 +0200
data backend: HDF5
generating machine: unknown
generating software: MALA (version: 1.1.0)
generating software dependencies: unknown

number of iterations: 1 (groupBased)
  all iterations: 0

number of meshes: 1
  all meshes:
    Density

number of particle species: 0

So /data/0/ is the openPMD iteration counter, which we use to name snapshots. Density/0 is one grid / array / Dataset (in hdf terms) / mesh (in openPMD terms) of shape 18 x 18 x 27. Multiple snapshots in one file would be called

/data/0/meshes/Density/0     Dataset {18, 18, 27}
/data/1/meshes/Density/0     Dataset {18, 18, 27}
/data/2/meshes/Density/0     Dataset {18, 18, 27}
...

Ba4O4

Contains DFT calculation output from a QuantumEspresso calculation for a barium oxide cell with 8 atoms, along with input scripts and pseudopotential to replicate this calculation. LDOS files are usually large, therefore this reduced example samples the LDOS somewhat inaccurately, in order to reduce storage size. The energy grid for the LDOS is 67 entries long, split along the energy axis at multiple points and sampled with a spacing of 0.5 eV. Only one snapshot is contained in the numpy format, all training and OpenPMD testing can be done with the beryllium data set. In detail, the following data files can be found:

File Name	Description
recreate_data/	Input scripts for QE
cubes/	.cube files for the local density of states
Ba_ncrs_v050_pbe.upf / O_ncrs_v050_pbe.upf	Pseudopotentials used for the QE calculation
BaO_snapshot0.out	Output file of QE. calculation
BaO_snapshot0.in.npy	Bispectrum descriptors numpy array
BaO_snapshot0.out.npy	Local density of states numpy array

Bispectrum descriptors of length 91 on 18 x 18 x 27 real space grid.

Note

In the last dimension of length 33, the first 3 entries are the grid coordinates / indices (an artifact of the SNAP vector generation). The actual features are snap_array[..., 3:].

>>> np.load('Be2/Be_snapshot1.in.npy').shape
(48, 48, 48, 33)

LDOS (67 / 64 points) on 48 x 48 x 48 real space grid.

Note

This is an LDOS split along the energy axis, meaning that upon loading within MALA, the values directly at the split are discarded. The energy axis is split in three places. Load with MALA to get the correct LDOS. The actual features have dimensionality 64!`.

>>> np.load('Be2/Be_snapshot1.out.npy').shape
(48, 48, 48, 67)