PaddleMaterials is an end-to-end AI4Materials toolkit built on the PaddlePaddle deep learning framework. Designed as a data-mechanism dual-driven platform for developing and deploying foundation models in materials science, PPMat enables researchers to efficiently build AI models and accelerate material discovery using pretrained models.
| Task | Description | Typical Applications |
|---|---|---|
| Property Prediction (PP) | Predict material properties from structure | Forward design or predict formation energy, band gap, elastic moduli etc. |
| Structure Generation (SG) | Generate novel crystal structures | Inverse design or structure generation |
| Machine Learning Interatomic Potential (MLIP) | Surrogate Model for DFT as ML potentials | Molecular dynamics simulations |
| Electronic Structure (ES) | Surrogate Model for DFT to predict physical field | Predict electronic density |
| Spectrum Elucidation (SE) | Reconstruct structures from spectra | NMR structure elucidation |
| Spectrum Enhancement (SPEN) | Enhance microscopy and spectrum signals | STEM image enhancement, denoising |
- Inorganic Crystals - Well-supported with multiple datasets and pretrained models
- Organic Molecules - Support for multiple datasets and pretrained models including small molecules and partial polymers
- ✅ Rich Pretrained Models & AI-ready Datasets - 50+ pretrained models ready for inference and Multiple curated datasets for training
- ✅ Multi-Task Integration - Unified framework across tasks of PP, SG, MLIP, ES, SE, SPEN etc.
- ✅ Multi-Hardware Support - Full support for NVIDIA GPUs and MetaX GPUs and Intel CPUs
- ✅ Production-Ready - Easy to use with standandlize design & distributed training, mixed precision, checkpoint recovery
| Task | Description | Link |
|---|---|---|
| Property Prediction (PP) | Predict formation energy, band gap, elastic properties | README |
| Structure Generation (SG) | Generate new crystal structures with diffusion models | README |
| Machine Learning Interatomic Potential (MLIP) | DFT-accurate potentials for molecular dynamics | README |
| Electronic Structure (ES) | Predict electronic structure properties | README |
| Spectrum Elucidation (SE) | Reconstruct molecular structures from NMR spectra | README |
| Spectrum Enhancement (SPEN) | Enhance microscopy and spectral signals | README |
| Task | Models | Dataset |
|---|---|---|
| Property Prediction | MEGNet, iComformer, DimeNet++, SphereNet | MP2018, MP2024, JARVIS, QM9, etc. |
| Structure Generation | MatterGen, DiffCSP | MP20, ALEX, etc. |
| Machine Learning Interatomic Potential | CHGNet, MatterSim, SphereNet | MPTRJ, MD17, etc. |
| Electronic Structure | InfGCN | QM9_ES, MP_ES, OMol25_MC_ES, etc. |
| Spectrum Elucidation | DiffNMR | MSD_NMR, etc. |
| Spectrum Enhancement | SFIN | SFIN-HAADF/BF, etc. |
Full model list: See MODEL_REGISTRY
Please refer to the installation document for your hardware environment. See SupportedHardwareList for more multi-hardware adaptation information.
Predict material formation energy using a pretrained MEGNet model:
python property_prediction/predict.py \
--model_name='megnet_mp2018_train_60k_e_form' \
--weights_name='best.pdparams' \
--input_format='cif' \
--input_path='./property_prediction/example_data/cifs/' \
--output_path='result_property_prediction/'Generate novel crystal structures using a pretrained MatterGen model:
python structure_generation/sample.py \
--model_name='mattergen_mp20' \
--weights_name='latest.pdparams' \
--output_path='result_mattergen_mp20/' \
--mode='by_num_atoms' \
--num_atoms=4Predict energy and forces using a pretrained MatterSim model:
python interatomic_potentials/predict.py \
--model_name='mattersim_1M' \
--weights_name='mattersim-v1.0.0-1M_model.pdparams' \
--input_format='cif' \
--input_path='./interatomic_potentials/example_data/cifs/' \
--output_path='result_interatomic_potentials/'Predict electron density from the bundled methane example using a pretrained InfGCN model:
python electronic_structure/predict.py \
--model_name='infgcn_qm9' \
--weights_name='best.pdparams' \
--input_format='mol' \
--input_path='electronic_structure/example_data/methane.mol' \
--grid_shape=8 \
--grid_batch_size=4096 \
--output_path='output/infgcn_qm9/methane'See the InfGCN prediction guide
for registered-model and local-checkpoint inference from MOL, CUBE, CHGCAR, and
density JSON inputs. Real test-split field examples are bundled under
electronic_structure/example_data/.
Run NMR spectrum elucidation using the bundled example and a pretrained DiffNMR model:
python spectrum_elucidation/sample.py \
--model_name='diffnmr_msdnmr_nless15' \
--weights_name='best.pdparams' \
--output_path='result_diffnmr_nless15/'Enhance STEM images using a pretrained SFIN model:
python spectrum_enhancement/predict.py \
--model_name='sfin_haadf_enhance' \
--weights_name='best.pdparams' \
--input_path='spectrum_enhancement/example_data/sfin_haadf.png' \
--output_path='result_sfin/'For training and fine-tuning, refer to the documentation.
Thanks to all contributors who have helped build PaddleMaterials！
Thanks for the following organiziton for cooprative support!
Join the PaddleMaterials WeChat group to discuss with us!
For developer, please refer to architecture.
PaddleMaterials is licensed under the Apache License 2.0.
@misc{paddlematerials2025,
title={PaddleMaterials, a deep learning toolkit based on PaddlePaddle for material science.},
author={PaddleMaterials Contributors},
howpublished = {\url{https://github.com/PaddlePaddle/PaddleMaterials}},
year={2025}
}This repository references code from the following projects:
PaddleScience | Matgl | CDVAE | DiffCSP | MatterGen | MatterSim | CHGNet | AIRS