Title: Accelerating Molecular Dynamics Through Speculative Sampling

Abstract: Molecular dynamics (MD) serves as a fundamental method for modeling the time-evolution of atomic systems. A significant limitation of MD, however, is its inherent serial nature, which complicates efforts to boost throughput for individual systems through parallel processing. To overcome this bottleneck, we present Langevin Speculative Dynamics (LSD), a distributed and model-agnostic speculative sampling technique designed to accelerate MD simulations without introducing additional relative error. Drawing inspiration from speculative approaches used in diffusion and language modeling, LSD employs a draft model to suggest rapid simulation steps, which are then validated in parallel against a slower, more accurate target model. This process utilizes a transport map to bridge the draft and target distributions. We adapt speculative sampling for second-order Langevin dynamics, mathematically deriving the potential speedup based on physical parameters. Our findings demonstrate that LSD offers a 3-9x speedup across various system and draft-target configurations. Furthermore, both theoretical analysis and empirical results confirm that LSD successfully generates trajectories consistent with the target model’s distribution.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC