GenSI/moma-modules

MoMa: Modular Deep Learning Framework for Material Property Prediction

📄 Paper | 🌐 Project Page | 💻 Code | 📊 Datasets

🚀 Overview

MoMa introduces a revolutionary modular approach to material property prediction that fundamentally shifts from traditional pre-training paradigms. Instead of using a single pre-trained model fine-tuned for each task, MoMa trains specialized modules across diverse material properties and adaptively composes them for downstream applications.

🎯 Key Achievements

• 🏆 State-of-the-art performance on 17 material property prediction datasets
• 📈 14% average improvement over strongest baselines
• 🔧 100+ specialized modules covering major material property domains
• ⚡ Superior few-shot and continual learning capabilities

🧠 Why MoMa?

Traditional Approach	MoMa Framework
Single pre-trained model → Fine-tune	Multiple specialized modules → Adaptive composition
Limited task specificity	Task-specific optimization
Poor generalization to new properties	Enhanced generalization through modularity
High computational cost for each task	Efficient module reuse and composition

📦 Available Modules (107 Total)

Our repository contains 107 pre-trained modules spanning 6 major domains of material science:

📊 Module Distribution by Domain

Domain	Count	Description
** Electronic Structure**	28	Band gaps, HOMO-LUMO, DOS, dielectric properties
** Thermodynamics**	20	Formation energy, Gibbs free energy, stability
** Spectroscopy**	24	EXAFS, XANES spectral features
** Mechanical**	8	Elastic moduli, piezoelectric properties
** Photovoltaic**	8	Solar cell performance metrics
** Adsorption**	8	Gas adsorption in MOFs
Thermoelectric	8	Seebeck coefficients, thermal conductivity
** Other Properties**	3	Specialized material characteristics

🔍 Featured Modules

🔬 Electronic Structure Modules

Module	Property	Description
HL_Gap.pt	HOMO-LUMO Gap	Electronic band gap prediction
HOMO_Energy.pt	HOMO Energy	Highest occupied molecular orbital energy
LUMO_Energy.pt	LUMO Energy	Lowest unoccupied molecular orbital energy
Polarizability.pt	Polarizability	Electron cloud deformation under external field
jarvis_bandgap.pt	Band Gap	Fundamental electronic property
mp_bandgap.pt	Material Band Gap	Electronic structure parameter

🌡️ Thermodynamics Modules

Module	Property	Description
jarvis_eform.pt	Formation Energy	Thermodynamic stability indicator
mp_eform.pt	Formation Energy	Energy of formation from elements
gibbs_free_energy.pt	Gibbs Free Energy	Chemical reaction spontaneity
surface_energy.pt	Surface Energy	Cost of creating new surfaces
mp_energy_above_hull.pt	Energy Above Hull	Phase stability metric

⚗️ Spectroscopy Modules

Module	Property	Description
spec_EXAFS_Fe_EdgeEnergy.pt	EXAFS Edge Energy	Iron K-edge absorption
spec_XANES_Co_PeakHeight.pt	XANES Peak Height	Cobalt absorption intensity
spec_EXAFS_Cu_WhiteLineHeight.pt	White Line Height	Copper spectral feature

📥 Download Individual Modules

import shutil
from huggingface_hub import hf_hub_download
# see all module names at https://huggingface.co/yuanhangtangle-air/moma-modules/blob/main/module-names.csv
# Download a specific module
cached_file = hf_hub_download(
    repo_id="yuanhangtangle-air/moma-modules",
    filename="Dipole_M.pt",  # Replace with desired module
    repo_type="model"
)

# Copy to your local directory
save_path = "./moma-hub/"
shutil.copy(cached_file, save_path)

📦 Download All Modules

import shutil
import pandas as pd
from huggingface_hub import hf_hub_download
from tqdm import tqdm

# fetch 
cached_file = hf_hub_download(
    repo_id="yuanhangtangle-air/moma-modules",
    filename="module-names.csv",
    repo_type="model"
)
df = pd.read_csv(cached_file)

MODULE_LIST = df['module_names']
cached_files = []
for module_name in tqdm(MODULE_LIST, desc="Downloading Modules"):
    cached_file = hf_hub_download(
    repo_id="yuanhangtangle-air/moma-modules",
        filename=module_name + ".pt",
        repo_type="model"
    )
    cached_files.append(cached_file)

# save to local path
save_path = "./moma-hub/"
for cached_file in cached_files:
    shutil.copy(cached_file, save_path)

📝 Citation

If you use MoMa in your research, please cite our paper:

@article{wang2025moma,
  title={MoMa: A Modular Deep Learning Framework for Material Property Prediction},
  author={Wang, Botian and Ouyang, Yawen and Li, Yaohui and Wang, Yiqun and Cui, Haorui and Zhang, Jianbing and Wang, Xiaonan and Ma, Wei-Ying and Zhou, Hao},
  journal={arXiv preprint arXiv:2502.15483},
  year={2025}
}

📜 License

This project is licensed under the MIT License.

🙏 Acknowledgments

We gratefully acknowledge the materials science community and the datasets that made this research possible. Special thanks to:

• Materials Project
• Open Quantum Materials Database
• Joint Automated Repository for Various Integrated Simulations
• And all other data contributors

---

🌟 Star this repository if you find it useful! 🌟

🚀 Accelerating Materials Discovery Through Modular AI 🚀