High-Dimensional Neural Network Potential for Liquid Electrolyte Simulations.

Area of Science:

• Materials Science
• Computational Chemistry
• Electrochemistry

Background:

• Liquid electrolytes are crucial for lithium-ion batteries, a key modern technology.
• Accurate computational tools for predicting electrolyte properties like viscosity and ionic diffusivity are lacking.
• Existing methods balance accuracy and efficiency, but a gap remains for first-principles simulations.

Purpose of the Study:

• To develop a machine learning-based force field for accurate and efficient simulation of liquid electrolytes.
• To bridge the accuracy gap between high-level quantum chemistry methods and classical force fields.
• To enable large-scale atomistic modeling of battery electrolytes.

Main Methods:

• Developed a machine learning force field using the QRNN deep neural network architecture.
• Incorporated long-range interactions and global charge equilibration into the model.
• Trained the model using non-periodic density functional theory (DFT) calculations (ωB97X-D3BJ/def2-TZVPD).

Main Results:

• The machine learning force field achieved quantitatively accurate predictions of material properties compared to experimental data.
• The model demonstrated efficiency comparable to classical force fields while maintaining high accuracy.
• Successfully applied to common carbonates and LiPF6, showing broad applicability.

Conclusions:

• The developed machine learning force field offers a promising approach for accurate atomistic simulations of liquid electrolytes.
• This methodology can be readily extended to various electrolyte systems and battery chemistries.
• Facilitates improved design and understanding of Li-ion battery materials.