Research map / Graph learning / GNN architectures
GFFMERGE: Efficient Merging of Graph Neural Force Fields and Beyond
Parth Verma, Parv Pratap Singh, Vipul Garg, Ishita Thakre, N M Anoop Krishnan, Sayan Ranu
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 |
|---|---|---|---|
| use closed-form embedding-alignment merging instead of weight averaging, Fisher, TIES or EMR | M3GNet force fields · MD17 5-Task Mix | Force MAE: 2.16 -> 0.89 | Improved |
| use closed-form embedding-alignment merging instead of baselines | Orb force fields · MD17 5-Task Mix | Force MAE: 3.98 -> 0.94 | Improved |
| merge LiPS20 solid-state models with baseline merging methods | M3GNet · LiPS20 | Force MAE: baselines up to 51.79 vs GFFMerge 0.72 (beta + gamma Li3PS4) | Hurt |
| add targeted fine-tuning of late interaction layers and force heads after closed-form merge (ablation) | GFFMerge · LiPS20 Li3P + Li2S | Force MAE | Improved |
| closed-form initialization before any fine-tuning (epoch 0) compared to baselines (ablation) | GFFMerge · held-out test set | Force MAE | Improved |
| fine-tune merged model with only 250 samples versus full data (ablation) | GFFMerge · few-shot fine-tuning | Force MAE | No clear effect |
| use GFFMerge instead of Joint FT for training | M3GNet and Orb · MD22 Ac-Ala3 + DHA + Stachyose | training time speedup: 7358.42 s -> 264.62 s (27.81x) | Improved |
| use GFFMerge instead of baselines in 1000-step MD rollouts | M3GNet · MD17 5-Task Mix | Force violation: 0.652 -> 0.562 (vs Wt Avg) | Improved |
| merge more domains simultaneously | baseline merging methods · MD17 | Force MAE | Hurt |
| use GNNMerge instead of Surgery for GCNs trained on disjoint label splits | GCN · in-domain node classification | accuracy: 78.01 -> 88.94 (average) | Improved |
| merge GCNs of different tasks (node classification and link prediction) with baselines | GCN · different tasks on different datasets | link prediction AUC | Hurt |
| merge MAG240M and Papers100M GraphSAGE models instead of joint training | GraphSAGE · OGB large scale | time and memory: 650GB RAM -> ~2 GB, under 8 minutes | Improved |
Closest papers by meaning
- On the Scalability of GNNs for Molecular Graphs (NeurIPS 2024 · similarity 0.87)
- Towards Fast, Specialized Machine Learning Force Fields: Distilling Foundation Models via Energy Hessians (ICLR 2025 · similarity 0.86)
- Neural P$^3$M: A Long-Range Interaction Modeling Enhancer for Geometric GNNs (NeurIPS 2024 · similarity 0.86)
- Neural Network Reparametrization for Accelerated Optimization in Molecular Simulations (NeurIPS 2024 · similarity 0.86)
- Pushing the Limits of All-Atom Geometric Graph Neural Networks: Pre-Training, Scaling, and Zero-Shot Transfer (ICLR 2025 · similarity 0.86)
- Long-Short-Range Message-Passing: A Physics-Informed Framework to Capture Non-Local Interaction for Scalable Molecular Dynamics Simulation (ICLR 2024 · similarity 0.85)
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.
