Self-learning multiscale simulation for achieving high accuracy and high efficiency simultaneously.

Biomolecular systems are inherently hierarchic and many simulation methods that try to integrate atomistic and coarse-grained (CG) models have been proposed, which are called multiscale simulations. Here, we propose a new multiscale molecular dynamics simulation method which can achieve high accuracy and high sampling efficiency simultaneously without aforehand knowledge on the CG potential and test it for a biomolecular system. In our method, a self-learning strategy is introduced to progressively improve the CG potential by an iterative way. (1) A CG model, coupled with the atomistic model, is used for obtaining CG structural ensemble, (2) which is mapped to the atomistic models. (3) The resulting atomistic ensemble is used for deriving the next-generation CG model. Two tests show that this method can rapidly improve the CG potential and achieve efficient sampling even starting from an unrealistic CG potential. The resulting free energy agreed well with the exact result and the convergence by the method was much faster than that by the replica exchange method. The method is generic and can be applied to many biological as well as nonbiological problems.