This directory contains the data and models of the paper "Stable and Accurate Orbital-Free Density Functional Theory Powered by Machine Learning".

It contains results from Kohn-Sham DFT calculations of molecules from the QM9 and QMUGS datasets, conducted at the PBE/6-31G(2df,p) level of theory with PySCF and partially with perturbed external potentials, yielding a more varied data distribution, as detailed in the associated paper. Furthermore, the results of density-fitting the Kohn-Sham data to yield labels in a linear combination of atomic basis functions (LCAB) Ansatz are included, with energies, electron density coefficients, and gradients w.r.t. the density. Finally, we provide checkpoints of models trained on said labels, which can be used to run orbital-free density optimization with STRUCTURES25.

Data extraction

The data is stored in tar files for smaller directories and split tar files for the larger directories. The data can be extracted by running the following line.

bash assemble_and_extract.sh

Set the environment variable DFT_DATA to the created data folder to run training or density optimization.

To delete the downloaded tar files after successful extraction, you can run

rm *.tar*

in this directory.

Data structure

After extraction, you will have a data folder which contains the following directories:

.
├── QM9_perturbed_fock
│   ├── dataset_statistics
│   ├── labels
│   ├── split.pkl
│   └── split.yaml
├── QMUGS
│   ├── QMUGSBin0.csv
│   ├── QMUGSLargeBins.csv
│   └── labels
├── QMUGSBin0QM9_perturbed_fock
│   ├── dataset_statistics
│   ├── split.pkl
│   └── split.yaml
├── QMUGSBin0_perturbed_fock
│   ├── dataset_statistics
│   ├── split.pkl
│   └── split.yaml
└── QMUGS_perturbed_fock
    └── labels

The labels directories contain the actual labels in the LCAB Ansatz, i.e., density coefficients, energies, and gradients of these energies w.r.t. the density. The structure is designed such that QM9_perturbed_fock, QMUGS, and QMUGS_perturbed_fock store all data, from raw to Kohn-Sham to labels. Other directories like QMUGSBin0QM9_perturbed_fock exist to create a split file and the corresponding dataset statistics.

Checkpoints

The checkpoints can be found in models/train/runs. To use them for density optimizations or fine-tunings, set DFT_MODELS to the models folder.

Code

The code can be accessed via zenodo (https://zenodo.org/records/14940433). It includes REPLICATION_GUIDE.md with information on how to replicate our results. The latest version of our functional will be provided at https://github.com/sciai-lab/structures25.