Skip to content

Perov-5: choose a model and generate

From nothing installed to candidate materials checked with a machine-learned interatomic potential (MLIP), with the commands that were run for this page and what they printed. Then: which of the public checkpoints to use for what.

Quick start · Which model · Your own targets · Your own data

Quick start

Three candidates for a direct band gap of 2 eV, generated and screened. Run on a laptop for this page: 29 seconds to generate, under 2 minutes to screen.

  1. Install.

    pip install "meidnet[stability] @ git+https://github.com/ABnano/MEIDNet.git"

    [stability] adds the MACE potential for step 3. Without it, steps 1 and 2 work the same.

  2. Generate. The demo uses the published model and the cubic halide perovskite family.

    meidnet demo --family halide --band-gap 2.0 --enthalpy -0.10 -n 3 --out runs/demo

    It printed:

    candidate predicted band gap (eV) predicted formation enthalpy (eV/atom) note
    NaNiI₃ 1.97 -0.10
    KMnI₃ 2.09 0.35
    NaSnI₃ 1.98 -1.31 extrapolating: outside the range of the training data

    Each candidate is a CIF file in runs/demo/generation/cifs/; runs/demo/generation_report.html shows every rule each one passed.

  3. Screen with the MLIP. MACE-MP-0 relaxes each crystal and computes its formation energy.

    meidnet screen runs/demo/generation/cifs --train-csv data/perov5/train.csv

    It printed:

    candidate MACE formation energy (eV/atom) stable unique novel
    NaNiI₃ -0.72 yes yes yes
    KMnI₃ -1.08 yes yes yes
    NaSnI₃ -0.94 yes yes yes

    3 of 3 are stable (formation energy at most 0.10 eV/atom), unique and novel. --train-csv needs meidnet download-data once; without it, novelty is skipped.

  4. Confirm. An MLIP is a first filter. Before any claim about a candidate, confirm its stability and its band gap with DFT.

The outputs of this run are kept in benchmarks/reproduction/perov5/demo/. No installation: the same search runs in the Studio.

Which model

All public checkpoints, measured under the same protocol (leaderboard): 54 candidates for three band-gap targets in three chemical families, screened with MACE-MP-0.

model delivered (of 54) stable, unique, novel (of 54) stable (share) band-gap target met, of candidates with a DFT value alignment (cosine) retrieval, top 1
MEIDNet (published model) 53 33 0.96 1 of 18 0.772 0.336
MEIDNet (shorter training) 52 30 0.92 1 of 18 0.480 0.297
MEIDNet (first early-fusion model) 54 35 0.94 3 of 16 0.655 0.028
MEIDNet (alignment training, seed 3) 47 33 0.87 2 of 8 0.954 ± 0.011 0.297 ± 0.062
Encoder screening (baseline) 54 46 0.85 – – –
Random sampling (baseline) 54 42 0.89 0 of 6 – –

Checkpoint files: MEIDNet (published model): dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth; MEIDNet (shorter training): …_propertyaware.pth; MEIDNet (first early-fusion model): dual_autoencoder_clip_earlyfusion.pth; MEIDNet (alignment training, seed 3): reproduction/meidnet_paper_rerun_seed3.pth. The alignment columns of the seed-3 row are the mean of the seven alignment models.

How to read this table, and what to start with

  • The checkpoints are close on stability. They deliver between 30 and 35 stable, unique and novel candidates of 54. Each is a single run; how much this count changes with the seed has not been measured.
  • Stable, unique and novel is not the design goal. The two baselines score higher on it (46 and 42 of 54), in part because they rarely return a material of the data set. Whether the band-gap target is met is only known where a DFT value exists: 8 to 18 candidates per model, too few to rank the models.
  • To generate, start with the published model: it is the model of meidnet demo and of the Studio, and its decoder (like that of the shorter training) was trained to rebuild a crystal from the property latent alone, which is what inverse design asks of it. Compare its candidates with the encoder screening of the same family (meidnet space), which needs no search.
  • To study the shared space, use the seven alignment models: their modalities agree most closely (cosine 0.95 against 0.77 for the published model), and seven seeds show how much a result depends on the seed. They were not trained to decode from properties alone: under this protocol the seed-3 model delivered 47 of the 54 candidates asked for, against 53 for the published model.

Your own targets

Targets, families, elements and rules are settings of one file, meidnet.yaml.

meidnet init --template perov5 -o meidnet.yaml       # a starting file
# edit generation.targets, generation.variant, generation.exclude_elements …
meidnet generate meidnet.yaml --model checkpoints/dual_autoencoder_clip_earlyfusion_propertyaware_2k.pth
meidnet screen runs/perov5/generation/cifs --train-csv data/perov5/train.csv

Change the target properties · Exclude or restrict elements · Change the material family · Add a rule

Your own data

A table with one row per material, a CIF for each row and one numeric column per property is enough to train your own model: Bring your own dataset. When you train:

  • keep the contrastive warm-up, train more than one seed and choose the checkpoint on the validation split (why);
  • expect the alignment of a short training to depend on the seed (figure);
  • hold materials out, and report the scores on them (what to expect).