Enhanced deep potential model for fast and accurate molecular dynamics: application to the hydrated electron

Area of Science:

• Computational Chemistry
• Materials Science
• Quantum Mechanics

Background:

• Neural network force fields offer a promising route to achieve high accuracy in molecular simulations at a reduced computational cost compared to traditional methods.
• Existing models face challenges in balancing accuracy, computational efficiency, and applicability to diverse systems, particularly for complex electronic phenomena.

Purpose of the Study:

• To introduce an enhanced Deep Potential network architecture for molecular simulations.
• To develop a lightweight and efficient neural network force field suitable for large-scale, accuracy-sensitive applications.
• To accurately model challenging systems like the solvated electron in water and investigate its properties.

Main Methods:

• Integration of a message-passing framework into the Deep Potential network architecture.
• Development of a new iterative model for Wannier center prediction to track electron positions.
• Application of the enhanced model to simulate the solvated electron in bulk water using extensive computational runs.

Main Results:

• The enhanced model achieves accuracy comparable to leading machine learning force fields with significant speed improvements.
• The model demonstrates transferability to larger systems, indicating its robustness.
• Simulations confirm the stability of the solvated electron's localized state within a cavity model.
• Accurate determination of structural and dynamical properties of the solvated electron in water.

Conclusions:

• The enhanced Deep Potential network provides a computationally efficient and accurate tool for molecular simulations.
• The developed model is well-suited for studying complex electronic systems, including the solvated electron.
• This work advances the capability of neural network force fields for large-scale, accuracy-demanding simulations.