Learn how to use OneScience for model training and development - Read Docs (Gitee) | GitHub Docs.

Project links: OneScience Gitee | OneScience GitHub | OneSkills Gitee | OneSkills GitHub

💧 MACE-water Usage Guidelines

Quick Start Guide

• OneScience Pre‑built Environment: The platform provides a ready‑to‑use full‑stack image. No manual environment setup is required – start model development immediately.
• Recommended Hyperparameters:
  • Cutoff radius r_max: typically 5.0–6.0 Å
  • Batch size batch_size: adjust according to GPU memory (e.g., 32 or 64)
  • Learning rate lr: 1e-3 – 1e-2
• Data Format: Extended XYZ files (.xyz) containing energies, atomic forces, and optionally virial stresses.
• Quick Single‑DCU Training: git clone https://gitee.com/hpccube/onescience.git cd onescience/examples/MaterialsChemistry/mace/ bash ./scripts/train.sh

1. Model Overview

MACE (Multi‑Atomic Cluster Expansion) is a state‑of‑the‑art machine learning force field that combines high‑order equivariant message‑passing neural networks (Equivariant MP‑GNNs) with the mathematical rigor of atomic cluster expansion (ACE). Its core features include:

• High‑accuracy predictions: Predict energies, atomic forces, and virial stresses with near‑DFT/xTB accuracy.
• Equivariant architecture: Built‑in rotational and translational equivariance provides strong physical expressiveness.
• Broad applicability: Suitable for molecular dynamics of small molecules as well as large‑scale simulations of crystalline materials – a powerful tool for materials design and screening.

The OneScience platform provides a ready‑to‑use full‑stack image:

✅ No manual environment installation required.
✅ Start model development and deployment immediately.

2. Environment Setup and Execution

(1) Underlying Build Process

The runtime environment is built on Python 3.11 and deeply optimized for Hygon DCU hardware. The core build steps are:

1. Load the DTK environment
   Load the appropriate DTK (Developer Tool Kit) via the cluster module system to provide low‑level GPU acceleration support.

2. Install PyTorch (DCU edition)
   Obtain the DCU‑optimized PyTorch wheel (.whl) from the Guanghe Developer Community to ensure perfect compatibility with the hardware instruction set.

3. Integrate OneScience services
   Install the core OneScience platform components from source to connect data pipelines and model interfaces.

(2) Build Script Example

1). Create and activate a dedicated Python 3.11 environment in Conda

conda create -n onescience_matchem python=3.11 -y
conda activate onescience_matchem
### 2). Load the underlying compiler and DTK environment (adjust module name as needed for your cluster)
   module load compiler/dtk/25.04.2
### 3). Install the DCU‑compatible PyTorch ecosystem (DAS 1.7 / DTK 25.04 / Python 3.10)
  Install core libraries: PyTorch, torchvision
pip install https://download.sourcefind.cn:65024/directlink/4/pytorch/DAS1.7/torch-2.5.1+das.opt1.dtk25042-cp310-cp310-manylinux_2_28_x86_64.whl
pip install https://download.sourcefind.cn:65024/directlink/4/vision/DAS1.7/torchvision-0.20.1+das.opt1.dtk25042-cp310-cp310-manylinux_2_28_x86_64.whl
### 4). Install OneScience framework and domain environment from source
git clone https://gitee.com/hpccube/onescience.git
cd onescience
pip install .

3. License

This code repository is licensed under the MIT License. The use of MACE‑water models is also subject to the MIT License. OneScience series (including Base and Chat) supports commercial use and distillation.

4. Citation

If you use MACE‑water or the OneScience platform in your research, please cite: @misc{onescience2025mace, title={MACE-water: A Multi-Atomic Cluster Expansion Force Field for Molecular Dynamics}, author={OneScience Team}, year={2025}, howpublished={\url{https://gitee.com/onescience-ai/onescience}}, note={Accessed: 2025-04-07} }

5. Contact

If you have any questions, please contact us at:

Email: xushl3@sugon.com