一张GPU跑10万原子！分子之心用AI把化学反应“拍”成了电影
Ranking
Observed public metrics from 1 member.
Merged summary
TL;DR - MolecularMind says its QuantaMind reactive machine-learning force field can simulate chemical reactions in systems approaching 100,000 atoms on one GPU, combining near-quantum accuracy with longer timescales and lower computational cost. The advance could make reaction-mechanism analysis and computer-guided protein, enzyme, and drug design more practical.
- QuantaMind was trained on key states from 5,286 reactions and uses mixed-fidelity data plus iterative high-accuracy calculations to improve weak spots.
- A Science Advances study simulated a 17,792-atom PETase system for 6 ns without prescribing the reaction pathway, capturing proton transfer, bond breaking and formation, and a complete catalytic cycle.
- Reported validation included agreement with experimental reaction energetics and correlations above 0.99 against quantum-mechanical calculations on sampled structures.
- The company reports scaling to roughly 100,000 atoms for hundreds of nanoseconds on a single GPU, with about 0.25 seconds per simulation step.
Sources (1)
一张GPU跑10万原子！分子之心用AI把化学反应“拍”成了电影
TL;DR - MolecularMind says its QuantaMind reactive machine-learning force field can simulate chemical reactions in systems approaching 100,000 atoms on one GPU, combining near-quantum accuracy with longer timescales and lower computational cost. The advance could make reaction-mechanism analysis and computer-guided protein, enzyme, and drug design more practical.
- QuantaMind was trained on key states from 5,286 reactions and uses mixed-fidelity data plus iterative high-accuracy calculations to improve weak spots.
- A Science Advances study simulated a 17,792-atom PETase system for 6 ns without prescribing the reaction pathway, capturing proton transfer, bond breaking and formation, and a complete catalytic cycle.
- Reported validation included agreement with experimental reaction energetics and correlations above 0.99 against quantum-mechanical calculations on sampled structures.
- The company reports scaling to roughly 100,000 atoms for hundreds of nanoseconds on a single GPU, with about 0.25 seconds per simulation step.