Research map / Probabilistic and scientific modeling / Molecules and proteins
Teaching Molecular Dynamics to a Non-Autoregressive Ionic Transport Predictor
Jiyeon Kim, Byungju Lee, Won-Yong Shin
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 model-level auxiliary modality learning (trajectory knowledge transfer) (ablation) | non-autoregressive predictor · Dataset 1 (Li MSD) 600K | MAE log10 MSD_Li: 0.395 -> 0.344 | Improved |
| Closed-form ridge initialization instead of gradient-based optimization (ablation) | predictor f1 · Dataset 1 1000K | MAE log10 MSD_Li: 0.455 -> 0.402 | Improved |
| Add structure-only regularization for dual-modal trainer (ablation) | dual-modal trainer g · Dataset 1 1200K | MAE log10 MSD_Li: 0.408 -> 0.390 | Improved |
| Use MOMENT embedding of trajectory spectrum (ablation) | trajectory embedding · Dataset 1 600K | MAE log10 MSD_Li: 0.429 -> 0.344 | Improved |
| Use SevenNet structure embedding instead of LiFlow (ablation) | structure embedding · Dataset 1 1000K | MAE log10 MSD_Li: 0.458 -> 0.402 | Improved |
| Add polynomial expansion to embeddings (ablation) | linear encoder · Dataset 1 800K | MAE log10 MSD_Li: 0.409 -> 0.367 | Improved |
| Add model-level AML (pretrain on Dataset 1) (ablation) | predictor f2 · Dataset 2 Na 1500K | MAE log10 D_Na: 0.146 -> 0.149 | No clear effect |
| Add model-level AML (ablation) | predictor f2 · Dataset 2 Na 1000K | MAE log10 D_Na: 0.201 -> 0.166 | Improved |
| Add both model- and data-level AML (ablation) | predictor f2 · Dataset 3 experimental | MAE log10 sigma_Li: 1.539 -> 1.388 | Improved |
| Initialize f2 encoder with W_x,T instead of W_trj (ablation) | predictor f2 · Dataset 2 Na 1000K | MAE log10 D_Na: 0.365 -> 0.166 | Improved |
| Non-autoregressive prediction instead of autoregressive LiFlow | Ours vs LiFlow · Dataset 1 | inference time (s): 2910 -> 14 | Improved |
| Use trajectory-aware training rather than plain non-autoregressive models | Ours vs Comformer · Dataset 1 1200K | MAE log10 MSD_Li: 0.760 -> 0.390 | Improved |
Closest papers by meaning
- FlashMD: long-stride, universal prediction of molecular dynamics (NeurIPS 2025 · similarity 0.88)
- All-atom Diffusion Transformers: Unified generative modelling of molecules and materials (ICML 2025 · similarity 0.86)
- Langevin-Informed Transfer Learning: Replacing the Target Samples by Black-Box Feedback (NeurIPS 2026 · similarity 0.86)
- Physics-Informed Weakly Supervised Learning For Interatomic Potentials (ICML 2025 · similarity 0.86)
- Learning Macroscopic Dynamics from Partial Microscopic Observations (NeurIPS 2024 · similarity 0.85)
- The dark side of the forces: assessing non-conservative force models for atomistic machine learning (ICML 2025 · 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.
