Reproduce “ATOM: A Pretrained Neural Operator for Multitask Molecular Dynamics” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Luke Thompson, Davy Guan, Dai Shi, Slade Matthews, Junbin Gao, Andi Han
A new pretrained neural operator, ATOM, accurately predicts molecular dynamics trajectories for unseen molecules and diverse timeframes, demonstrating robust zero-shot generalization. The model reduces State-to-Trajectory Mean Squared Error by nearly 40% in out-of-domain settings compared to previous methods like EGNO and introduces the TG80 dataset for multitask pretraining.
Key insights
- A quasi-equivariant architecture, combining initial E(3)-equivariant lifting with unconstrained transformer blocks, offers a beneficial balance between physical symmetry awareness and model expressivity.
- Non-autoregressive operator learning with heterogeneous temporal attention and T-RoPE facilitates efficient, parallel trajectory prediction and robust handling of irregular time lags.
- Multitask pretraining across diverse molecular systems and timeframes on the TG80 dataset enables robust zero-shot generalization to unseen compounds and varying time horizons.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.