Perov-5 — the paper's experiment

The published MEIDNet model: cubic ABX₃ perovskites from the Perov-5 dataset (CDVAE split of Castelli et al.), properties formation enthalpy (heat_all, eV/atom) and direct band gap (dir_gap, eV).

Use the published checkpoint

meidnet demo --family halide --band-gap 2.0            # or: oxide, chalcogenide, nitride
meidnet studio                                          # explore it live

The three checkpoints from the paper are in checkpoints/; ..._propertyaware_2k.pth is the recommended one. meidnet info checkpoints/<file> describes any of them. They load into MEIDNet 2 unchanged.

Retrain from scratch

meidnet download-data                       # train/val/test CSVs → data/perov5/ (18,928 rows, 18 MB)
meidnet train examples/perov5/meidnet.yaml  # 200 epochs ≈ 1 h on a laptop GPU
meidnet generate examples/perov5/meidnet.yaml

examples/perov5/meidnet.yaml reproduces the paper's settings, including two that differ from MEIDNet 2's defaults: normalize: false for both properties and align_to_prototype: false (atoms in file order), and the 1200-epoch alignment ramp.

The paper's generated structures

examples/perov5/paper_results/ holds the 26 CIFs (halide, oxide, chalcogenide) reported in the paper. meidnet screen examples/perov5/paper_results --train-csv data/perov5/train.csv recomputes their MACE stability.

Regression against v1

tests/legacy_v1/ contains the v1.0 code byte for byte. pytest -m slow runs it and MEIDNet 2 on the same targets and asserts identical CIFs, predictions and file names.