Skip to content

DATASET.md

Latest commit

 

History

History
194 lines (160 loc) · 8.63 KB

DATASET.md

File metadata and controls

194 lines (160 loc) · 8.63 KB

Materials Discovery: GNoME Datasets

GNoME

The GNoME dataset provides ~381,000 novel structures that update the convex hull of known stable materials. Associated files are provided in the public Cloud Bucket defined by gs://gdm_materials_discovery with the following organization:

gdm_materials_discovery
└───gnome_data
│   │   stable_materials_summary.csv
│   │   stable_materials_r2scan.csv
│   │   by_composition.zip
│   │   by_id.zip
│   │   by_reduced_formula.zip
│   └───auxiliary_gnome_data
|       |   a2c_supporting_data.json
└───external_data
|   |   mp_snapshot_summary.csv
|   |   external_materials_summary.csv

There are two main options for downloading the dataset to a local directory.

  • Command line: the command line interface can be used to directly download structures from the public bucket via a command of the form gcloud storage cp --recursive gs://gdm_materials_discovery/ data/. The gcloud storage command can be installed following the Google Cloud CLI install instructions.

  • Python script: Helper scripts have also been provided, which can be run using either scripts/download_data_wget.py or scripts/download_data_cloud.py. The latter may be preferable if the user already has the Google Cloud CLI authorized.

    First, install the required dependencies. It is best to do this inside a virtual environment:

    python -m venv ~/venv/gnome
source ~/venv/gnome/bin/activate
pip install absl-py google-cloud-storage

    The google-cloud-storage package is only required for the download_data_cloud.py script. Then run either:

    python scripts/download_data_wget.py

    or, after following the Google Cloud CLI install instructions,

    gcloud auth application-default login
python scripts/download_data_gcloud.py

    Both scripts take an optional flag, --data_dir, to control the directory the data is downloaded to. By default, the data directory in the current working directory is used, and created if required.

Convex Hull

First, in stable_materials_summary.csv, we provide a CSV containing the compositions of all novel materials as well as corresponding energies (along with a number of other properties). This file is the simplest and most useful for calculating the convex hull energies over all stable materials. New materials can be benchmarked against this file for decomposition energy estimates. We have updated this csv to ensure minimal overlap with newer versions of external datasets such as OQMD and MP, based on snapshots in June 2024.

Note, construction of the complete convex hull requires energies from Materials Project (MP), the Open Quantum Materials Database (OQMD), and WBM. We have included external_materials_summary.csv to provide composition and associated convex hull entries with elemental compositions that match these external datasets, though in some cases the improved energies correspond to lower energy structures. To ensure compatibility between DFT calculations, we correct the energies of all entries in the Materials Project that contain any of Ga, Ge, Li, Mg, Na, elements for which the pseudopotentials we use and the ones used by MP differ. The correction per atom is provided below:

  pp_corr = {"Ga": -0.0028805, "Ge": 0.10417085, "Li": -0.00301278, "Mg": 0.0924014, "Na": -0.00447437}

The combination of these two datasets provides the updated convex hull and can be used for evaluating the stability of other computational experiments. In additional, for the example colabs, we also provide a snapshot of the stable crystals from the Materials Project, enabling visualization of the exploration spaces of GNoME.

Properties

Beyond energy, a number of additional properties may be of interest for the novel crystal structures. stable_materials_summary.csv provides a number of additional descriptors that may be helpful when processing or looking for crystals with specific properties.

A list of the properties provided and textual descriptions is provided below:

  • Composition: alphabetically-ordered composition
  • MaterialId: a unique id corresponding to the entry
  • Reduced Formula: reduced chemical formula
  • Elements: chemical system
  • NSites: number of atoms
  • Volume: volume in units Å^3
  • Density: density in units Å^3 / atom
  • Point Group: assigned point group
  • Space Group: assigned space group
  • Space Group Number: assigned space group number
  • Crystal System: assigned crystal system
  • UnCorrected Energy: uncorrected energy
  • Corrected Energy: energy adjusted by MP2020 corrections
  • Formation Energy Per Atom: normalized energy corrected by reference elements
  • Decomposition Energy Per Atom: decomposition energy relative to the downloaded Materials Project convex hull
  • Dimensionality Cheon: dimensionality predicted by Cheon et al. 2017
  • Bandgap: calculated bandgap (if available)
  • Is Train: in training set for associated machine learning models
  • Decomposition Energy Per Atom All: distance to convex hull of all entries
  • Decomposition Energy Per Atom Relative: distance to convex hull of all entries except for the current
  • Decomposition Energy Per Atom MP: distance to convex hull of all entries from Materials Project
  • Decomposition Energy Per Atom MP OQMD: distance to convex hull of all entries from Materials Project + Open Quantum Materials Database (including recalculations)

Formation energies per atom are all calculated with respect to a fixed set of elemental references from the Materials Project (and not lowered if we found a lower) energy elemental structure. For clarity, we have provided a list of the elemental references used in JSON form in elemental_references.json

Structures

Each of the associated rows of the CSV provided above is associated with a crystal structure. We provide compressed directories of associated CIFs, a standard file format within the materials science community. Note, three versions of the compressed directories exist, where file names allow for lookup by unique identifier, reduced formula, or by composition.

r²SCAN

Validation of the associated structures was completed using r²SCAN. stable_materials_r2scan.csv provides all r²SCAN calculations performed on stable materials. Note, stability metrics change with the choice functional (as discussed in the associated paper), so not all released materials remain stable according to this metric.

Caveats

Due to numerical precision (and errors arising from the computational simulations), we use a threshold of 5e-5 eV as the threshold for determining whether a material is on the convex hull. For all measurements, the provided materials update the convex hull of a snapshotted version of Materials Project and similar databases. Therefore, as more crystals are discovered by the scientific community, the above set may not remain stable.

Versioning

Below, we keeps notes about any upgrades made to the dataset as well as approximate timing.

  • (11/29/23) Initial dataset release
  • (12/1/23) Re-introduce paper filters to remove un-physical energies; add 2 missing columns ('Dimensionality Cheon' and 'Is Train')
  • (8/21/24) Adjust threshold to 1meV/atom off the hull significantly increasing the number of released crystals, update to use "require_bound" version of MP corrections, re-relax structure where VASP did not update lattice coordinates, provide convex hull entries from recomputations of MP / OQMD / ..., provided colabs for accessing, fixed file naming in by_id.zip
  • (11/21/24) Ensure consistency in formation energies, provide used elemental references + pseudopotential corrections used in the GNoME paper

Disclaimer

Additional data that are not as part of the core GNoME effort (such as e.g. band gaps) are provided as-is, at PBE-level, are work in progress, and may be updated over time. Continued efforts to characterize the electronic properties of stable materials may update / correct these values. More accurate calculations and processing are a work-in-progress. Other auxiliary data related to GNoME efforts (e.g. structure prediction with a2c) is available under auxiliary_gnome_data.