Scalable neural network driven molecular dynamics simulation

Molecular dynamics (MD) simulation is an essential tool for condensed matter physics, materials science, structural/mechanistic biology, and multi-agent systems. Despite their successes, traditional numerical integration methods for solving Hamilton’s equations of motion are computationally intensive, limiting simulations to short time scales. Recent advancements in machine learning have opened new avenues for accelerating MD simulations. This work introduces the Local Update Function network (LUFnet), a transformer-based neural network designed to increase time integration step sizes significantly while maintaining simulation stability and accuracy. LUFnet integrates local spatial and temporal information, enabling efficient rollout for long-time-scale simulations. By preserving key symmetries such as translational invariance, Galilean coordinate transformation invariance, and particle exchange symmetry, LUFnet achieves robust performance across different thermodynamic states. LUFnet is designed to accommodate general physical models (e.g., Lennard-Jones, Coulomb potential, on lattice systems). Its framework allows the model to be trained on small systems and directly applied to larger systems, maintaining computation efficiency and memory usage that scales linearly with the number of particles. Benchmarked on Lennard-Jones systems, LUFnet demonstrated minimal accuracy degradation even after rollout over many large time integration steps, offering an effective approach to molecular dynamics simulations.