UBio-MolFM: Quantum-Accurate Simulation of Million-Atom Systems
量子位 · wechat · 2026-08-26
UBio Innovation Institute released UBio-MolFMv1.5, a machine-learned interatomic potential function aiming to resolve the trade-off between accuracy and scale in molecular dynamics simulations. The model provides quantum chemistry accuracy at classical force field speeds without re-parameterization for new systems. In real-world tests on the KcsA potassium channel and Cyclosporine A conformational free energy, UBio-MolFM revealed qualitative misjudgments possible with classical force fields. Technically, it uses the E2Former-V2 architecture with hybrid cutoff radius stacking and a three-stage curriculum learning, achieving an effective receptive field of 23Å. Performance wise, it reduces force errors by 40%-45% on biomolecules compared to baselines, supporting 180k atoms on a single card and up to 1.44 million atoms on an 8-card node. Model weights, code, and the technical report are now fully open source.
More from Research
- Game2World Engine removes gameplay UI for better world model training — GameWorldData · 2026-08-26
- Skild AI unveils S1: robot foundation model that learns 10-minute novel tasks from one video — deepakpathak · 2026-08-26
- Composable Curricula and difficulty pool for optimized training — samsja19 · 2026-08-26
- UAI2026 announces keynote speakers in causal learning and responsible AI — LesterMackey · 2026-08-26
- A 30-question deep dive into embeddings, vector search and retrieval — techNmak · 2026-08-26
- Deepak Pathak Demonstrates In-Context Learning for Robotics with One-Shot Video Prompts — deepakpathak · 2026-08-26