Mahpe/polyanion-pretrained-mlips

Polyanion Pretrained MLIPs (PaiNN / cPaiNN / CHGNet / MACE)

Twenty-three machine-learning interatomic potentials trained on the polyanion sodium cathode DFT dataset (298,144 structures): system-specific PaiNN/cPaiNN models plus fine-tuned universal foundation models (CHGNet, MACE-MP-0).

Paper: Limitations of Foundation Models in Energy Materials Simulations: A Case Study in Polyanion Sodium Cathode Materials — Advanced Intelligent Discovery (2026).

Dataset: Scientific Data (2025) · Hugging Face · DTU archive

Code: dtu-energy/cPaiNN

Model inventory

Folder	Architecture	Models	Targets
Pretrained_models_PaiNN/	PaiNN	8	energy, forces, stress
Pretrained_models_cPaiNN/	cPaiNN	8	energy, forces, stress, magmom, Bader charge
Pretrained_models_ensemble_cPaiNN/	cPaiNN ensemble	3	energy, forces, stress, magmom, Bader charge
Pretrained_universal_models/CHGNet/	CHGNet	2	energy, forces, stress, magmom, charge
Pretrained_universal_models/Mace_MP_0_large/	MACE-MP-0	2	energy, forces

Each PaiNN/cPaiNN model directory contains:

• best_model.pth — validation-best checkpoint
• arguments.json — hyperparameters
• datasplits.json — train/validation split indices
• commandline_args.txt — launch command

Hidden dimensions (node_size) and interaction depth (num_interactions) are encoded in the folder name, e.g. Polyanion_ionicsteps_magmom_bader_512_3 → 512 nodes, 3 interaction layers, with magmom + Bader charge heads.

Universal foundation models (Pretrained_universal_models/) include CHGNet and MACE-MP-0 (large) checkpoints, each trained from scratch and fine-tuned on the polyanion dataset for comparison in AIDi (2026).

Quick start

git lfs install
git clone https://huggingface.co/Mahpe/polyanion-pretrained-mlips
pip install git+https://github.com/dtu-energy/cPaiNN.git

import torch
from cpainn import CPaiNN  # see cPaiNN repo for exact API

checkpoint = torch.load(
    "Pretrained_models_cPaiNN/Polyanion_ionicsteps_magmom_bader_512_3/best_model.pth",
    map_location="cpu",
)
# Load into CPaiNN using arguments.json hyperparameters — see cPaiNN train.py / inference examples

Training & benchmark figures

Validation MAE curves below are extracted from the original printlog.txt training logs bundled with each checkpoint.

Benchmark figures from the related AIDi / electronic-entropy study on NaFePO₄ charge ordering (arXiv:2603.26471):

Figure 3. GA structures and relative energies — cPaiNN vs MACE vs CHGNet vs DFT.

Figure 6. Na_xFePO₄ convex hulls before/after electronic-entropy embedding (MAE vs DFT).

Citation

If you use these models, please cite the dataset and cPaiNN papers:

@article{petersen2026foundation,
  title={Limitations of Foundation Models in Energy Materials Simulations: A Case Study in Polyanion Sodium Cathode Materials},
  author={Hoffmann Petersen, Martin and others},
  journal={Advanced Intelligent Discovery},
  year={2026},
  doi={10.1002/aidi.202500065}
}

@article{petersen2025polyanion,
  title={Dataset exploring the atomic scale structure and ionic dynamics of polyanion sodium cathode materials},
  author={Hoffmann Petersen, Martin and others},
  journal={Scientific Data},
  year={2025},
  doi={10.1038/s41597-025-05799-8}
}