Research map / Label-efficient and robust learning / Self-supervised and contrastive
TriForces: Augmenting Atomistic GNNs for Transferable Representations
Ali Ramlaoui, Alexandre Duval, Hannah Bull, Victor Schmidt, Hugues Talbot, Fragkiskos D. Malliaros, Joseph Musielewicz
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 |
|---|---|---|---|
| add composition and structure streams plus SSL pretraining (TriForces) | Orb-v3 Conservative · OMat24 val | energy MAE (meV/atom): 107 -> 19.4 | Improved |
| add SSL pretraining on top of the three streams (ablation) | Orb-v3 Conservative · OMat24 val | force MAE (meV/A): 149 -> 95.5 | Improved |
| add three streams without SSL (ablation) | Orb-v3 Conservative · OMat24 val | force MAE (meV/A): 150 -> 149 | No clear effect |
| add three streams without SSL (ablation) | MACE · OMat24 val | energy MAE (meV/atom): 117 -> 81.6 | Improved |
| add three streams without SSL (ablation) | eSEN · MPtrj subset | force MAE (meV/A): 75.9 -> 93.3 | Hurt |
| add SSL pretraining on top of streams (ablation) | Orb-v3 Direct · OMat24 val | stress MAE: 6.5 -> 7.8 | Hurt |
| add TriForces three-stream plus SSL at 20K training samples | eSEN Direct · OMat24 fine-tuning | energy MAE (meV/atom): 81.3 -> 34.6 | Improved |
| add SSL initialization at full OMat24 scale (ablation) | TriForces-Streams (eSEN) · Full OMat24 | energy MAE (meV/atom): 13.7 -> 13.5 | No clear effect |
| use SSL instead of supervised pretraining on LeMat-Bulk (ablation) | eSEN · OMat24 fine-tuning | force MAE (meV/A): 106.0 -> 86.8 | Improved |
| add composition stream to interaction stream (ablation) | eSEN-sm conservative · OMat24-2M (no SSL) | energy MAE (meV/atom): 102 -> 16 | Improved |
| add structure stream to composition+interaction (ablation) | eSEN-sm conservative · OMat24-2M (no SSL) | crystal-system probe accuracy (%): 50.6 -> 68.1 | Improved |
| add TriForces to Orb-v3 on Matbench Discovery | Orb-v3 · Matbench Discovery (unique prototypes) | MAE (eV/atom): 0.024 to 0.020 | Improved |
Closest papers by meaning
- Physics-Informed Weakly Supervised Learning For Interatomic Potentials (ICML 2025 · similarity 0.87)
- The Importance of Being Scalable: Improving the Speed and Accuracy of Neural Network Interatomic Potentials Across Chemical Domains (NeurIPS 2024 · similarity 0.87)
- FlashTP: Fused, Sparsity-Aware Tensor Product for Machine Learning Interatomic Potentials (ICML 2025 · similarity 0.87)
- MLIP Arena: Advancing Fairness and Transparency in Machine Learning Interatomic Potentials via an Open, Accessible Benchmark Platform (NeurIPS 2025 · similarity 0.86)
- Towards Fast, Specialized Machine Learning Force Fields: Distilling Foundation Models via Energy Hessians (ICLR 2025 · similarity 0.86)
- Benchmarking Compositional Generalisation for Machine Learning Interatomic Potentials (NeurIPS 2026 · 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.
