Adaptive coupling of a deep neural network potential to a classical force field

Area of Science:

• Computational chemistry
• Materials science
• Chemical physics

Background:

• Molecular simulations face a trade-off between accuracy and computational efficiency.
• Accurate methods like ab initio molecular dynamics are computationally expensive.
• Classical force fields offer efficiency but lack accuracy for complex phenomena.

Purpose of the Study:

• To introduce an adaptive modeling method (AMM) that integrates deep neural network potentials with classical force fields.
• To address the accuracy-efficiency dilemma in molecular simulations.
• To enable accurate modeling of critical system regions while maintaining computational efficiency.

Main Methods:

• The AMM decomposes the simulated system into three regions: high-accuracy (Deep Potential Molecular Dynamics - DeePMD), low-accuracy (classical force field), and a transition region.
• The DeePMD model is trained using ab initio molecular dynamics data.
• A force interpolation scheme and thermodynamics force are applied in the transition region for seamless integration.

Main Results:

• The developed AMM successfully integrates DeePMD with classical force fields.
• A smooth transition between high-accuracy and efficient regions is achieved.
• The method demonstrates feasibility and promise in a liquid water system simulation.

Conclusions:

• The AMM offers a viable solution to the accuracy-efficiency challenge in molecular simulations.
• This hybrid approach allows for targeted high-accuracy modeling where needed.
• The method shows potential for broader applications in simulating complex molecular systems.