MoMa: Modular Learning for Materials

Introduction

Material property prediction is critical for accelerating materials discovery, yet material tasks exhibit significant diversity (spanning electronic, thermal, mechanical properties across crystals, molecules, and more) and disparity (governed by distinct physical laws). Naively pre-train then fine-tune often fails to address these challenges, where a monolithic model suffers from negative transfer.

MoMa (Modular learning for Materials) reframes learning as modular composition. Instead of forcing all tasks into one shared model, MoMa trains specialized modules on diverse high-resource tasks and adaptively composes synergistic modules tailored to each downstream scenario, enabling effective knowledge reuse while mitigating interference.

Highlights

14% average improvement over the strongest baseline across 17 material property prediction tasks
Best in 16 out of 17 tasks, spanning electronic, thermal, mechanical, and phonon-related properties
18 open-source modules released on Hugging Face for community reuse
Superior few-shot performance under realistic label scarcity

Framework

MoMa operates in two stages: (1) Module Training & Centralization, which distills transferable knowledge from diverse high-resource tasks into a shared MoMa Hub; and (2) Adaptive Module Composition (AMC) & Fine-tuning, where MoMa selects and weights modules via a training-free kNN estimator and convex optimization, then fine-tunes the composed model.

The MoMa framework. (a) Module Training & Centralization: train specialized modules and store them in MoMa Hub. (b) Adaptive Module Composition: estimate module relevance via kNN, optimize composition weights, and fine-tune.

Key Findings

Beyond strong empirical performance, our analysis reveals three insights about modular material learning:

📈

Scaling Law of Modules

MoMa improves monotonically as more modules are added to the hub, showing strong and predictable scaling behavior.

More modules = better performance

🎯

Few-Shot Superiority

MoMa's advantage becomes more pronounced in few-shot settings, highlighting data efficiency under label scarcity.

Greatest gains where labeled data is scarce

🔍

Interpretable Insights

The learned composition weights reveal meaningful relationships between material properties.

Composition weights encode transferable material knowledge

Results

Evaluation across 17 downstream material property prediction tasks. All results are the average of five data splits (MAE).

Main results for 17 material property prediction tasks. MoMa (Full) achieves the best average rank of 1.35, outperforming all baselines with an average improvement of 14%.

Citation

If you find this work useful, please cite our paper:

@inproceedings{wang2026moma, title={MoMa: A Simple Modular Learning Framework for Material Property Prediction}, author={Wang, Botian and Ouyang, Yawen and Li, Yaohui and Pan, Mianzhi and Cui, Haorui and Wang, Yiqun and Zhang, Jianbing and Wang, Xiaonan and Ma, Wei-Ying and Zhou, Hao and others}, booktitle={International Conference on Learning Representations}, volume={2026}, pages={45830--45855}, year={2026} }