oib12/nmdl-mattergen-e

MatterGen-e-

Diffusion models that generate an inorganic crystal structure and a per-atom electronic descriptor at the same time: one reverse-diffusion run produces the lattice, the atomic positions, the atomic numbers and either the atom-projected density of states (ADOS) or the Bader charge of every atom.

Code: nmdl-mizo/mattergen-e- · Paper: npj Comput Mater 12, 332 (2026) · Built on MatterGen; not affiliated with Microsoft.

Models

Directory	Jointly generated	Per-atom dim	Training set	Conditioning
atomic_dos_joint_model	ADOS latent	30	mp-20-curated	—
atomic_dos_joint_model_finetuned	ADOS latent	30	mp-20-curated	formation energy, energy above hull, band gap
atomic_dos_joint_model_mp_20	ADOS latent	30	mp-20	—
bader_charge_joint_model	Bader charge	1	mp-20-curated	—
bader_charge_joint_model_finetuned	Bader charge	1	mp-20-curated	formation energy, energy above hull, band gap

The 30 numbers an ADOS model writes per atom are not a spectrum — they are the latent vector of an autoencoder. Decode them with the matching checkpoint from oib12/nmdl-mattergen-e_atomic-dos-autoencoder; the wrong one gives a plausible-looking but meaningless DOS.

For DOS generation, atomic_dos_joint_model_mp_20 is the recommended model; atomic_dos_joint_model is the one behind the results in the paper.

See the code repository for how to load these checkpoints and generate structures.

Training data

Both datasets are built from MP-20 — Materials Project (MP, CC BY 4.0) structures with at most 20 atoms per unit cell.

• mp-20-curated — the MP-20 structures for which MP publishes both the charge density and the DOS. ADOS comes from the published DOS; the Bader charge was computed for this work by Bader analysis of the published charge density.
• mp-20 — MP-20 with the DOS recomputed from scratch by our own DFT calculations. ADOS only.

ADOS is stored as a 30-dimensional latent encoding the DOS over E − E_F ∈ [−15, 5] eV on 512 points.

Please credit the Materials Project alongside the citation below.

Limitations

• No unit cells beyond 20 atoms, no organic or non-crystalline materials.
• The generated descriptors are model predictions, not a substitute for DFT.

Citation

Okuda, I., Takahara, I. & Mizoguchi, T. Inverse materials design via joint generation of crystal structures and local electronic descriptors. npj Comput Mater 12, 332 (2026). https://doi.org/10.1038/s41524-026-02322-4

@article{Okuda2026MatterGenE,
  title   = {Inverse materials design via joint generation of crystal structures and local electronic descriptors},
  author  = {Okuda, Ibuki and Takahara, Izumi and Mizoguchi, Teruyasu},
  journal = {npj Computational Materials},
  volume  = {12},
  pages   = {332},
  year    = {2026},
  doi     = {10.1038/s41524-026-02322-4},
  url     = {https://doi.org/10.1038/s41524-026-02322-4},
}

MatterGen-e- builds on MatterGen, so please also cite the original work: Zeni et al., A generative model for inorganic materials design, Nature (2025), doi:10.1038/s41586-025-08628-5.