STRUCTURES25

Machine-learned orbital-free density functional theory

Pretrained models for optimizing molecular electron densities and energies, developed by SciAI Lab, Heidelberg University. Equivariant graph neural networks learn the kinetic-plus-exchange-correlation energy functional from reference DFT data.

GitHub · Documentation · Paper · Data

Available models

Model	Training molecules	CLI name
QM9	QM9	str25_qm9
QMugs	Small-molecule QMugs subset	str25_qmugs

Both checkpoints support H, C, N, O, and F and use training data augmented with perturbed external potentials. Intended for molecular OF-DFT research within the chemical scope of the training data; accuracy and convergence on new systems require validation.

Get started

Follow the installation and setup guide to install mldft, download the models, and configure their locations. Then run:

mldft example.xyz --model str25_qm9

Use --model str25_qmugs for the QMugs checkpoint. See the usage guide for options and the replication guide for technical details and benchmarks.

Reference

Remme et al., Stable and Accurate Orbital-Free Density Functional Theory Powered by Machine Learning, J. Am. Chem. Soc. 147, 28851–28859 (2025). DOI · BibTeX