jhaens/solidacid-0_model

SolidAcid-0 MLIPs

Machine-learned interatomic potentials (MLIPs) for solid acid proton conductors, built on the MACE architecture.

This repository shares two families of models:

1. Local-universal (base) models: trained on the full SolidAcid-0 dataset spanning several solid-acid chemistries, intended as general-purpose potentials across the solid acid compositional space.
2. Domain-specific (transfer-learned) models: the base models fine-tuned on individual compounds for higher accuracy within a single chemistry.

The training/fine-tuning datasets are hosted separately on the Hugging Face Hub: https://huggingface.co/datasets/jhaens/solidacid-0_model_dataset

---

Repository layout

.
├── tiny/                       # base local-universal model, size "tiny"
│   ├── config_train.yaml
│   └── tiny_seed{1111..5555}.model
├── small/                      # base local-universal model, size "small"
│   ├── config_train.yaml
│   └── small_seed{1111..5555}.model
├── medium/                     # base local-universal model, size "medium"
│   ├── config_train.yaml
│   └── medium_seed{1111..3333}.model
├── tiny_transferlearned/       # domain-specific fine-tunes of the tiny base model
│   └── csh2po4/
│       ├── config_finetune.yaml
│       └── tiny_transferlearned_csh2po4_seed{1111..5555}.model
└── small_transferlearned/      # domain-specific fine-tunes of the small base model
    ├── csh2po4/
    ├── csh2aso4/
    ├── cshseo4/
    └── cshso4/
        ├── config_finetune.yaml
        └── small_transferlearned_<system>_seed{1111..5555}.model

Each model is provided as a seed ensemble. Different seeds are independently initialized/trained models with identical architecture and data; use them together to estimate committee (ensemble) uncertainty, or pick a single seed for production runs.

---

Model sizes

All base models are MACE potentials with the same core hyperparameters (see config_train.yaml in each directory). The sizes differ in their equivariant feature content:

Size	hidden_irreps	max_ell	Seeds
tiny	64x0e	2	1111–5555
small	128x0e	3	1111–5555
medium	128x0e + 128x1o	3	1111–3333

Shared architecture / training settings (base models):

• model: MACE, default_dtype: float64
• r_max: 6.0, correlation: 3, num_interactions: 2
• interaction_first: RealAgnosticDensityInteractionBlock, interaction: RealAgnosticDensityResidualInteractionBlock
• pair_repulsion: True, distance_transform: Agnesi
• loss: universal, forces_weight: 10, energy_weight: 1.0 (no stress)
• EMA (ema_decay: 0.999), amsgrad, weight_decay: 1e-8, clip_grad: 10.0
• lr: 0.005, batch_size: 8, max_num_epochs: 25

Transfer-learned (domain-specific) models

The domain-specific models are produced by fine-tuning a single base seed (<size>_seed1111.model) on per-compound data (see each config_finetune.yaml):

• Fine-tuned for max_num_epochs: 30, lr: 0.005 (0.01 for cshso4)
• multiheads_finetuning: false (single-head naive/transfer fine-tuning)
• Per-system atomic reference energies (E0s) and reference keys (REF_TotEnergy, REF_Force) as recorded in the config
• Released as seed ensembles (1111–5555)

---

Usage

These are standard MACE model files and can be used with the mace-torch package via ASE.

Installation

pip install mace-torch
# optional:
pip install cuequivariance

Single model with ASE

from ase.io import read
from mace.calculators import MACECalculator

atoms = read("your_structure.xyz")

calc = MACECalculator(
    model_paths="small/small_seed1111.model",
    device="cuda",              # or "cpu"
    default_dtype="float64",
)
atoms.calc = calc

print(atoms.get_potential_energy())
print(atoms.get_forces())

Ensemble (committee) for uncertainty estimates

Pass all seeds of a given model to average predictions and obtain a per-atom force committee spread:

from mace.calculators import MACECalculator

seeds = [1111, 2222, 3333, 4444, 5555]
calc = MACECalculator(
    model_paths=[f"small/small_seed{s}.model" for s in seeds],
    device="cuda",
    default_dtype="float64",
)

---

Data

Training and fine-tuning datasets: https://huggingface.co/datasets/jhaens/solidacid-0_model_dataset

The config_train.yaml / config_finetune.yaml files reference the internal split paths (train/val/test) used at training time; the corresponding data is provided in the dataset repository above.

---

Citation

If you use these models, please cite the corresponding preprint and the MACE publication:

@unpublished{solidacid0,
      title={ABC},
      author={H{\"a}nseroth, Jonas and von Stackelberg, Rose and Dre{\ss}ler, Christian},
      archivePrefix={arXiv},
      year={2026},
      eprint={2609.xxx}
}

@inproceedings{Batatia2022mace,
      title={{MACE}: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields},
      author={Ilyes Batatia and David Peter Kovacs and Gregor N. C. Simm and Christoph Ortner and Gabor Csanyi},
      booktitle={Advances in Neural Information Processing Systems},
      editor={Alice H. Oh and Alekh Agarwal and Danielle Belgrave and Kyunghyun Cho},
      year={2022},
      url={https://openreview.net/forum?id=YPpSngE-ZU}
}

License

The MLIP models are published and distributed under the MIT License.