Iterative Neural Variational Solver of Committor Function in Under-damped Langevin Dynamics
This work develops an iterative neural variational solver to compute the committor function for under-damped Langevin dynamics, which is vital for accurately modeling rare transitions in real systems across different friction levels. By decomposing the system's dynamics and using a generalized Concus-Golub-Widlund splitting method with neural networks, the method provides a practical tool to resolve sharp transition boundaries.