Research map / Probabilistic and scientific modeling / Molecules and proteins
MoMa: A Simple Modular Learning Framework for Material Property Prediction
Botian Wang, Yawen Ouyang, Yaohui Li, Mianzhi Pan, yuanhang tang, Haorui Cui, Yiqun Wang, Jianbing Zhang, Xiaonan Wang, Wei-Ying Ma, Hao Zhou
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| compose modules with MoMa instead of fine-tuning JMP directly | JMP · Matminer 17 tasks | MAE: average improvement of 14.0% over JMP-FT, better on 14 tasks | Improved |
| keep modules separate and compose adaptively instead of multi-task pretraining on all hub tasks | JMP · Matminer 17 tasks | MAE: 24.8% average margin over JMP-MT, better on 16 of 17 tasks | Improved |
| multi-task pretraining on hub tasks (JMP-MT) versus plain fine-tuning (ablation) | JMP · Matminer 17 tasks | average rank: 3.12 -> 4.53 | Hurt |
| use MoMa on the Orb (GNS) backbone | Orb-v2 · Matminer 17 tasks | MAE: 6.1% average boost, better on 13/17 tasks | Improved |
| AMC weighted selection instead of uniform average of selected modules (ablation) | MoMa · Matminer 17 tasks | test MAE: 11.0% average MAE increase without weights | Improved |
| AMC module selection instead of averaging all modules (Model Soup) (ablation) | MoMa · Matminer 17 tasks | test MAE: 18.0% average MAE increase for All Average | Improved |
| AMC selection instead of random module selection (ablation) | MoMa · Matminer 17 tasks | test MAE: 20.2% average MAE increase for Random Selection | Improved |
| AMC instead of LoRAHub black-box search for composition (ablation) | MoMa · Matminer 17 tasks | test MAE: 21.8% average MAE reduction, better on 15 tasks | Improved |
| AMC instead of a learned routing network over 18 modules (JMP-(18)) (ablation) | MoMa · Matminer 17 tasks | test MAE: 15.5% average MAE reduction, better on 17 tasks | Improved |
| AMC convex optimization instead of Softmax Weighting heuristic (ablation) | MoMa · Matminer 17 tasks | test MAE: 13.7% average MAE reduction, better on 12 tasks | Improved |
| use MoMa in 10-shot setting | JMP · Few-shot 10-shot | average normalized test MAE: 0.7003 -> 0.5503 | Improved |
| scale hub from 5 to 30 modules (adding 12 QM9 modules) (ablation) | MoMa · Matminer 17 tasks | average normalized test MAE: 0.2040 -> 0.1759 | Improved |
Closest papers by meaning
- A Denoising Pre-training Framework for Accelerating Novel Material Discovery (AAAI 2025 · similarity 0.88)
- Offline Materials Optimization with CliqueFlowmer (NeurIPS 2026 · similarity 0.87)
- MatExpert: Decomposing Materials Discovery By Mimicking Human Experts (ICLR 2025 · similarity 0.86)
- Crystalformer: Infinitely Connected Attention for Periodic Structure Encoding (ICLR 2024 · similarity 0.86)
- A Diffusion-Based Pre-training Framework for Crystal Property Prediction (AAAI 2024 · similarity 0.86)
- Generative Hierarchical Materials Search (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
