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

Area of Science:

• Computational chemistry
• Biophysics
• Molecular dynamics simulations

Background:

• Biomolecular systems exhibit hierarchical structures, necessitating advanced simulation techniques.
• Existing multiscale simulation methods often require prior knowledge of coarse-grained potentials.
• Integrating atomistic and coarse-grained (CG) models presents challenges in accuracy and efficiency.

Purpose of the Study:

• To develop a novel multiscale molecular dynamics simulation method.
• To achieve high accuracy and sampling efficiency simultaneously.
• To eliminate the need for a priori knowledge of the coarse-grained potential.

Main Methods:

• A self-learning strategy iteratively refines the coarse-grained (CG) potential.
• A CG model coupled with an atomistic model generates a CG structural ensemble.
• The atomistic ensemble derived from the CG model is used to create the next-generation CG model.

Main Results:

• The proposed method rapidly improves the CG potential, even from an unrealistic starting point.
• Efficient sampling is achieved, surpassing the performance of replica exchange methods.
• The calculated free energy shows excellent agreement with exact results.

Conclusions:

• This generic multiscale simulation method offers a powerful approach for biomolecular and non-biological systems.
• The self-learning strategy enables accurate and efficient simulations without prior potential knowledge.
• The method demonstrates significantly faster convergence compared to traditional techniques.