Force-free molecular dynamics for fast and accurate long-timescale simulations

Molecular dynamics (MD) simulations are a cornerstone of scientific research, providing critical insights into material properties and molecular behavior. However, their high computational cost often limits the accessible timescales and system sizes. While many data-driven approaches aim to accelerate force evaluations, they remain constrained by the need for small integration time steps. In this work, we introduce a transferable and data-efficient framework based on autoregressive equivariant message-passing networks that directly updates atomic positions and velocities, lifting traditional numerical integration constraints. We validate our approach across diverse systems—including small molecules, crystalline materials, and bulk liquids—demonstrating strong agreement with reference MD simulations for structural, dynamical, and energetic properties. For all systems, our method enables time step extensions of at least one order of magnitude compared to conventional MD simulations. By efficiently generating trajcetories of large systems over extended timescales, our framework has the potential to accelerate materials discovery and could help revealing physical phenomena beyond the scope of traditional methods.