oib12/nmdl-mattergen-e_atomic-dos-autoencoder

MatterGen-e- ADOS autoencoder

1D convolutional autoencoders that compress the atom-projected density of states (ADOS) of a single atom into a 30-dimensional latent vector and back. They define the latent space that the MatterGen-e- ADOS models generate, so one of these checkpoints is needed to turn the 30 numbers those models write per atom into an actual spectrum.

Code: nmdl-mizo/mattergen-e- · Paper: npj Comput Mater 12, 332 (2026)

Checkpoints

Plain PyTorch state_dict files; the architecture lives in the code repository (mattergen.decode_ados.autoencoder).

File	Trained on	Pairs with
mp_20_curated_-15eV_to_5eV_30dim.pt	DOS published by MP (mp-20-curated)	atomic_dos_joint_model, atomic_dos_joint_model_finetuned
mp_20_-15eV_to_5eV_30dim.pt	DOS recomputed by our own DFT (mp-20)	atomic_dos_joint_model_mp_20

A latent space belongs to one autoencoder. Decoding with the wrong checkpoint gives a plausible-looking but meaningless DOS, and nothing in the files makes the mismatch detectable.

Representation

DOS on 512 points covering E − E_F ∈ [−15, 5] eV, spin-summed and projected onto one atomic site. The encoder applies log1p before the convolution stack and ends in a scaled tanh, so every training latent lies in (−1, 1); the decoder inverts the transform with Softplus then expm1, making the reconstruction non-negative by construction.

The energy axis is numpy.linspace(-15.0, 5.0, 512). It is not stored in the checkpoint — only the file name records it — so the caller has to supply it.

Training data

Atom-projected densities of states of MP-20 structures from the Materials Project (MP, CC BY 4.0). mp-20-curated is the MP-20 structures for which MP publishes both the charge density and the DOS, and takes the published DOS; mp-20 is MP-20 with the DOS recomputed from scratch by our own DFT calculations. Please credit the Materials Project alongside the citation below.

Limitations

• Exponential tail. Softplus → expm1 makes the decoder exponential in its pre-activation, so a latent slightly outside the training distribution decodes to a DOS orders of magnitude above anything physical — 11 of 23,838 generated atoms in one run. Filter those out before averaging or integrating.
• 512 points compressed into 30 numbers: fine structure is not recoverable.
• Spin-summed, atom-projected total DOS only — not spin- or orbital-resolved.

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},
}