MEIDNet — pretrained Perov-5 models

MEIDNet (Multimodal Equivariant Inverse Design Network) designs crystalline materials from the properties you want. One shared latent space holds crystal structures and their properties; a prototype material family with chemistry rules defines what may be generated; a latent search proposes candidates that pass every rule and sit closest to the target.

Try it now	MEIDNet Prism — live Studio (nothing to install; bring your own table and train in the browser)
Paper	A. Babu, R. A. Gouvêa, P. Vandergheynst, G.-M. Rignanese, npj Computational Materials (2026) — doi:10.1038/s41524-026-02153-3 · arXiv:2601.22009
Code	github.com/ABnano/MEIDNet (MIT) · documentation · Colab notebooks
Benchmarks	MEIDNet Benchmarks — Perov-5 leaderboards (inverse design, property prediction, representation) with baselines; contribute yours

Files

file	what it is
dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth	the production model of the paper: early fusion, property-aware decoding, 2000 epochs with a contrastive warm-up over the first 1200. Use this one.
dual_autoencoder_clip_earlyfusion_propertyaware.pth	the same architecture, shorter training
dual_autoencoder_clip_earlyfusion.pth	the earliest ablation (no property-aware decoding)
meidnet.yaml	the configuration that reproduces the Perov-5 experiment with the MEIDNet 2 package
perovskite_abx3.yaml	the cubic ABX₃ family file: prototype sites, allowed elements per site, oxidation states, rules

Each checkpoint is 2.8 MB (about 0.7 M parameters) and runs on a laptop CPU.

What the model does

• Inputs: a crystal structure (CIF, up to 20 atoms per cell) and/or scalar properties — here the direct band gap (dir_gap, eV) and the formation enthalpy (heat_all, eV/atom).
• Model: an equivariant graph encoder for the structure and an MLP encoder for the properties are aligned contrastively (CLIP-style) into one 128-dimensional latent space; the joint latent is the average of the two (early fusion); decoders reconstruct the crystal and the properties.
• Inverse design: start at the latent of the target properties, optimise a population of latents, decode each into one element per prototype site, keep the candidates that pass every chemistry rule, rank by closeness to the target. Candidates must be confirmed by DFT or experiment; the package ships a MACE-based stability / uniqueness / novelty screen.

Training data: Perov-5 (CDVAE split; 11,356 training structures). Element coverage follows that data: oxides, nitrides, fluorides, sulfides and their mixtures. Predictions for elements absent from it (for example Cl, Br, I and most lanthanides) are extrapolations — the Studio and the reports say so.

Use it

pip install git+https://github.com/ABnano/MEIDNet.git
meidnet demo                         # downloads this checkpoint and designs three candidates

from huggingface_hub import hf_hub_download
from meidnet.checkpoint import load_checkpoint, describe

path = hf_hub_download("Babu09/MEIDNet", "dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth")
lm = load_checkpoint(path)
print(describe(lm))                  # properties, units, training ranges

With a configuration file (meidnet.yaml from this repository, model_path pointing at the checkpoint):

meidnet generate meidnet.yaml --quick      # candidates + a plain-language HTML report
meidnet studio meidnet.yaml                # the same workflow as a live web page

Results reported in the paper (Perov-5)

quantity	value
cosine similarity between the structure and property latents of the same material	≈ 0.97
L2 distance between those latents	≈ 0.24
inverse-design campaign: candidates generated → stable, unique and novel	140 → 19 (13.6 %)

These numbers are quoted from the paper. The Perov-5 benchmark evaluates this checkpoint with the public code under a fixed protocol, next to baselines: inverse design (stable, unique and novel candidates), property prediction and representation on the test split.

Citation

@article{meidnet2026,
  title   = {MEIDNet: Multimodal generative AI framework for inverse materials design},
  author  = {Anand Babu and Rog{\'e}rio Almeida Gouv{\^e}a and Pierre Vandergheynst and Gian-Marco Rignanese},
  journal = {npj Computational Materials},
  year    = {2026},
  doi     = {10.1038/s41524-026-02153-3}
}

Software: Anand Babu, MEIDNet (MIT), https://github.com/ABnano/MEIDNet.

Further reading

• A. Babu, N. M. A. Krishnan, Multimodal and cross-modal learning techniques, APL Machine Learning 4, 030901 (2026). doi:10.1063/5.0346744
• A. Babu, R. Almeida Gouvêa, G.-M. Rignanese, Toward automated discovery with generative models multimodal learning and closed loop workflows in inverse materials design, Cell Reports Physical Science 7, 103561 (2026). doi:10.1016/j.xcrp.2026.103561