Accelerating Molecular Dynamics Simulations with Foundation Neural Network Models Using Multiple Time Steps and...

Area of Science:

• Computational Chemistry
• Molecular Dynamics Simulations
• Machine Learning in Science

Background:

• Molecular dynamics (MD) simulations are crucial for understanding molecular behavior.
• Neural network potentials (NNPs) offer high accuracy but are computationally expensive.
• Accelerating MD simulations is essential for tackling larger and more complex systems.

Purpose of the Study:

• To introduce a novel Distilled Multi-Time-Step (DMTS) strategy for accelerating MD simulations.
• To leverage foundation neural network models for enhanced simulation efficiency.
• To maintain the accuracy of simulations while significantly reducing computational cost.

Main Methods:

• Developed a dual-level neural network architecture for MD simulations.
• Employed a distillation process to create a faster, lower-fidelity model from an accurate NNP.
• Integrated the distilled model within a Reversible Reference System Propagator Algorithm (RESPA)-like framework.
• Utilized active learning to enhance simulation stability, particularly for solvated proteins.

Main Results:

• Achieved significant speedups in MD simulations: nearly 4-fold for homogeneous systems and 3-fold for large solvated proteins.
• Demonstrated that the distilled model (3.5 Å cutoff) accurately captures fast-varying forces, primarily bonded interactions.
• Preserved both static and dynamic properties of the simulated systems, confirming the approach's accuracy.
• Enabled evaluation of the costly NNP every 3-6 fs, a substantial increase from the standard 1 fs timestep.

Conclusions:

• The DMTS strategy effectively accelerates molecular dynamics simulations using neural network potentials.
• This approach maintains high accuracy, comparable to standard methods, while offering substantial performance gains.
• DMTS reduces the computational performance gap between neural network potentials and classical force fields.
• The strategy is versatile and applicable to various neural network potentials and molecular systems.