Accelerated Langevin Dynamics Simulation via Neural Network-Driven Importance Sampling
Year: 2026
Type: preprint
Abstract: Atomistic simulations are often restricted by timescale limitations when systems become trapped in metastable energy basins. The transitions between these metastable states are rare events and they dictate the long-term evolution of the system. We present an importance sampling framework to accelerate the time scale of Langevin dynamics simulations. The framework uses a neural network parameterized importance function to bias the dynamics, enhancing the efficiency of rare-transition sampling while preserving relative probabilities between transition paths. We provide a rigorous mathematical formulation to recover the original transition rates between metastable states from the biased dynamics, and use a branching random walk algorithm to control the statistical variance of the estimated rates. We validate the framework on 2- and 14-dimensional problems, as well as the system of 7 Lennard-Jones discs (d = 14) model. We further apply it to two realistic materials systems: the diffusion of a Pt adatom on the Pt(001) surface (d = 387) and the migration of a vacancy in BCC \alpha-iron (d = 2055). The framework provides a scalable foundation for accelerating the simulation of rare events in atomistic systems.
Source: ChemRxiv
Authors: Michael Kim, Wei Cai
Institution: Stanford University
Language: English
Subfield: Physical and Theoretical Chemistry
Field: Chemistry
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7221108039