Research map / Architectures and efficiency / Pruning and efficient tuning
Sparsity-promoting Fine-tuning for Equivariant Materials Foundation Model
Youngwoo Cho, Seunghoon Yi, Wooil Yang, Sungmo Kang, Young-Woo Son, Jaegul Choo, Joonseok Lee, Soo Kyung Kim, Hongkee Yoon
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 sparsity-promoting fine-tuning (L) instead of full fine-tuning on LAM Sn energy | MACE-MP-0b3 · LAM Sn | energy MAE (meV/atom): 32.74 -> 2.18 | Improved |
| use sparse adaptation instead of ELoRA low-rank adaptation on TM-O-Spin forces | MACE-MP-0b3 with spin layers · TM-O-Spin | force MAE (meV/A): 116.31 -> 70.46 | Improved |
| use sparse adaptation (L) instead of full fine-tuning on TM-O-Spin energy | MACE-MP-0b3 with spin layers · TM-O-Spin | energy MAE (meV/atom): 11.58 -> 9.50 | Improved |
| use sparse adaptation (L) instead of ELoRA on rMD17 Aspirin forces | MACE-OFF23 · rMD17 Aspirin | force MAE (meV/A): 8.52 -> 7.56 | Improved |
| switch from low sparsity (lambda_tau 0.01) to high sparsity (0.3) on rMD17 Aspirin forces (ablation) | sparse fine-tuning · rMD17 Aspirin | force MAE (meV/A): 7.56 -> 8.22 | Hurt |
| update only the FCTP layers instead of the linear layers (ablation) | sparse fine-tuning · Ti (LAM) | energy and force MAE | Hurt |
| update both linear and FCTP layers (All) instead of linear only (ablation) | sparse fine-tuning · Ti (LAM) | energy and force MAE | No clear effect |
| use Linear-only vs All updates, comparing module sparsity growth with lambda_tau (ablation) | sparse fine-tuning · Ti (LAM) | layer-wise sparsity | Mixed |
| use high-sparsity setting (lambda_tau 0.3) vs low on TM-O-Spin energy (ablation) | sparse fine-tuning · TM-O-Spin | energy MAE (meV/atom): 9.50 -> 10.57 | Hurt |
| use sparse fine-tuning instead of full fine-tuning on large MP-mag dataset forces | MACE-MP-0b3 with spin layers · MP-mag | force MAE (meV/A): 37.56 -> 42.11 | Hurt |
| use sparse All (lambda_tau 0.05) instead of full fine-tuning on MP-mag energy | MACE-MP-0b3 with spin layers · MP-mag | energy MAE (meV/atom): 18.75 -> 16.56 | Improved |
| fine-tune with sparse method on NequIP-OAM-L instead of full fine-tuning (Aspirin energy) | NequIP-OAM-L · rMD17 Aspirin | energy MAE: 0.37 -> 0.33 | Improved |
Closest papers by meaning
- ELoRA: Low-Rank Adaptation for Equivariant GNNs (ICML 2025 · similarity 0.90)
- Learning the Electronic Hamiltonian of Large Atomic Structures (ICML 2025 · similarity 0.87)
- Efficient and Scalable Density Functional Theory Hamiltonian Prediction through Adaptive Sparsity (ICML 2025 · similarity 0.87)
- Higher-Rank Irreducible Cartesian Tensors for Equivariant Message Passing (NeurIPS 2024 · similarity 0.87)
- Learning Smooth and Expressive Interatomic Potentials for Physical Property Prediction (ICML 2025 · similarity 0.86)
- Physics-Informed Weakly Supervised Learning For Interatomic Potentials (ICML 2025 · 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.
