Mdlvzip

: The encoder takes the complex atomic positions and maps them into a tiny "bottleneck" size.

: Only the compressed latent data and the trained model weights need to be stored.

: You can share compact "latent representations" instead of entire raw trajectories, making collaboration easier. Mdlvzip

: It shrinks data by more than 95% compared to raw formats.

: The decoder reconstructs the atomic positions with high geometric accuracy. : The encoder takes the complex atomic positions

: Since small geometric shifts can affect energy readings, a short energy minimization step is often used after reconstruction to restore physical realism. Getting Started

The core of MDZip is a neural network called a . Mdlvzip

: Despite high compression, it preserves ensemble-level features like RMSD fluctuations and radius of gyration.