Faster Molecular Dynamics with Neural Network Potentials via Distilled Multiple Time-Stepping and Nonconservative...

Area of Science:

• Computational Chemistry
• Materials Science
• Artificial Intelligence in Science

Background:

• Atomistic molecular dynamics (MD) simulations are crucial for understanding molecular behavior.
• Accelerating MD simulations is essential for tackling larger and more complex systems.
• Foundation neural network models offer promise for accurate molecular simulations.

Purpose of the Study:

• To introduce the Distilled Multi-Time-Step (DMTS) approach with nonconservative (NC) forces for accelerated MD simulations.
• To enhance the efficiency and stability of simulations using neural network potentials.
• To demonstrate the applicability and performance of the DMTS-NC scheme.

Main Methods:

• Development of the DMTS-NC strategy, employing a dual-level Reversible Reference System Propagator Algorithm (RESPA) formalism.
• Coupling an accurate conservative potential with a simplified distilled representation for nonconservative forces.
• Incorporation of physical priors like equivariance and force cancellation into the distilled architecture.
• Combination with hydrogen mass repartitioning (HMR) and High Hydrogen Friction (HHF) for extended time steps.

Main Results:

• The DMTS-NC scheme achieves 15-30% greater speedup compared to its conservative counterpart.
• The approach demonstrates excellent agreement with force data, limiting discrepancies between models.
• Extended time steps up to 10 fs are maintained with stability and accuracy.
• Speedups ranging from 3.66 to 5.64 were observed when distilling MACE-OFF23.

Conclusions:

• The DMTS-NC approach offers a robust, stable, and efficient method for accelerating atomistic MD simulations.
• It requires no fine-tuning, simplifying implementation and maximizing efficiency.
• The method is versatile, applicable to any neural network potential (NNPs) and computationally intensive approaches.