Score Dynamics: Scaling Molecular Dynamics with Picoseconds Time Steps via Conditional Diffusion Model.

Area of Science:

• Computational Chemistry
• Molecular Dynamics
• Machine Learning

Background:

• Molecular Dynamics (MD) simulations are crucial for understanding molecular behavior but are computationally expensive.
• Existing methods often require small timesteps, limiting simulation speed and scope.
• Accelerating MD simulations is essential for exploring larger systems and longer timescales.

Purpose of the Study:

• To introduce Score Dynamics (SD), a novel framework for learning accelerated evolution operators.
• To enable large timestep simulations from Molecular Dynamics (MD) data.
• To develop efficient computational models for molecular systems.

Main Methods:

• Developed Score Dynamics (SD), a framework centered around scores (derivatives of transition log-probability).
• Utilized graph neural networks to construct SD models for molecular systems.
• Validated SD using alanine dipeptide and short alkanes in aqueous solution, with 10 ps timesteps.

Main Results:

• SD models accurately predict equilibrium properties and kinetic rates/paths, matching MD simulation results.
• The SD implementation achieves a speedup of approximately two orders of magnitude compared to traditional MD.
• Demonstrated the framework's efficacy on realistic molecular systems.

Conclusions:

• Score Dynamics (SD) provides a general and efficient framework for accelerating molecular dynamics simulations.
• SD successfully learns large timesteps, offering significant computational speedups.
• Future work will focus on addressing open challenges to further improve SD performance.