[NeurIPS 2026 Oral] Compact SO(3) Equivariant Atomistic Foundation Models via Structural Pruning

Paper

This repository implements structural pruning for SO(3) equivariant atomistic foundation models. The method removes complete irreducible representation blocks along the channel and order dimensions, preserving equivariance. The compact models can be retrained on a subset of the pre-training data and fine-tuned for downstream tasks.

We provide the pruning scripts and retraining configurations used in our paper for MACE-MP, MACE-OFF, SevenNet, and eSCN.

Setup

git clone https://github.com/HPC-AI-Team/SO3-Pruning.git
cd SO3-Pruning

Run all commands below from the repository root. Replace /path/to/... with your checkpoint, data, and output paths.

Use a separate environment for each model library. The scripts have been checked with the following package versions:

Model	Package	Official repository
MACE-MP and MACE-OFF	mace-torch==0.3.14	MACE
SevenNet	sevenn==0.12.0	SevenNet
eSCN	fairchem-core==1.10.0	FAIR-Chem

Install the package for the model you want to use in its environment:

# MACE environment
python -m pip install mace-torch==0.3.14 tqdm

# SevenNet environment
python -m pip install sevenn==0.12.0

# eSCN environment
python -m pip install fairchem-core==1.10.0

Follow the corresponding library's installation instructions for PyTorch, CUDA, and compiled dependencies.

MACE

Source checkpoints

L denotes the maximum hidden-feature order, and C denotes the channel count per order.

Model	Source checkpoint	Source	Supported targets
MACE-MP	MACE_MPtrj_2022.9.model	L1 C256	L0–L1, C1–C256
MACE-MP	2024-01-07-mace-128-L2_epoch-199.model	L2 C128	L0–L2, C1–C128
MACE-OFF23	MACE-OFF23_large.model	L2 C224	L0–L2, C1–C224

Download the source checkpoints from the MACE-MP release and the MACE-OFF23 repository.

1. Compute channel importance scores

Calibration is needed only when reducing the channel count. Provide a representative structure subset in an ASE-readable format, such as extended XYZ:

python MACE/pruning/calibration.py \
    --model /path/to/MACE_MPtrj_2022.9.model \
    --data /path/to/calibration.xyz \
    --output /path/to/mp_scores.pt \
    --device cuda

The script computes Taylor channel importance scores and displays calibration progress. Use --device cpu for CPU calibration. For MACE-OFF, run the same script with the OFF source checkpoint and molecular calibration structures.

Use scores computed from the same source checkpoint that you will prune.

2. Prune the model

For combined order and channel pruning of MACE-MP, for example L1 C256 to L0 C128:

python MACE/pruning/prune_mp.py \
    --model /path/to/MACE_MPtrj_2022.9.model \
    --scores /path/to/mp_scores.pt \
    --target-l 0 \
    --target-channels 128 \
    --output /path/to/mp_pruned.model

For order-only pruning, keep all source channels and omit --scores:

python MACE/pruning/prune_mp.py \
    --model /path/to/2024-01-07-mace-128-L2_epoch-199.model \
    --target-l 0 \
    --target-channels 128 \
    --output /path/to/mp_l0.model

For MACE-OFF23 Large to L1 C128:

python MACE/pruning/prune_off.py \
    --model /path/to/MACE-OFF23_large.model \
    --scores /path/to/off_scores.pt \
    --target-l 1 \
    --target-channels 128 \
    --output /path/to/off_pruned.model

3. Retrain

Set the checkpoint, dataset, and output paths in train_mp.sh or train_off.sh, then run:

bash MACE/train_mp.sh

or:

bash MACE/train_off.sh

Training settings are in mp.yaml and off.yaml. The scripts load the pruned model through MACE's --foundation_model option.

SevenNet

The pruning script converts SevenNet-l3i5 to an L2 model with five interaction layers and hidden channels 128x0e+64x1e+32x2e.

Download checkpoint_l3i5.pth, then run:

python SevenNet/pruning/prune.py \
    --model /path/to/checkpoint_l3i5.pth \
    --output /path/to/sevennet_pruned.pth

This fixed L3-to-L2 conversion does not require calibration scores.

For retraining, set train.continue.checkpoint, data.load_trainset_path, and data.load_validset_path in SevenNet-0.yaml. Prepare the structure data using SevenNet's data tools. Set the output directory in train.sh, then run:

bash SevenNet/train.sh

The example script runs on one GPU. The configured batch size is 1024. For distributed training, divide this global batch size by the number of GPUs when setting the per-GPU batch size.

eSCN

Download the eSCN-L6-M2-Lay12-S2EF-OC20-All+MD checkpoint from the official model list.

For L6-to-L4 pruning:

python eSCN/pruning/prune.py \
    --model /path/to/escn_l6_m2_lay12_all_md_s2ef.pt \
    --target-l 4 \
    --output /path/to/escn_pruned.pt

The script reduces lmax while retaining mmax, channels, and layer count. It supports single-resolution eSCN checkpoints. The target must satisfy mmax <= target_l < source_l, so the L6 M2 source supports target L2–L5. Calibration scores are not needed.

For retraining, prepare the OC20 S2EF data in FAIR-Chem's LMDB format. Set the data paths in train_1m.yml or train_2m.yml, and set the checkpoint and output paths in train.sh:

bash eSCN/train.sh

The script uses 8 GPUs and selects train_1m.yml by default. This configuration samples one million structures from the 2M training split. Select train_2m.yml to use the full split. Both configurations use L4, so update model.lmax_list if you prune to another order.

Acknowledgments

This implementation builds on MACE, e3nn, SevenNet, and FAIR-Chem.

Citation

If you find this work useful, please consider citing our paper:

@inproceedings{wang2026compact,
  title     = {Compact {SO(3)} Equivariant Atomistic Foundation Models via Structural Pruning},
  author    = {Wang, Chen and Hu, Siyu and Tan, Guangming and Jia, Weile},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}