Dielectric Saturation in Water from a Long-Range Machine Learning Model.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning in physics

Background:

• Neural network potentials offer high accuracy but traditionally lack long-range interaction modeling.
• Accurate simulation of polar liquids requires accounting for electrostatic interactions.
• Existing models often neglect long-range effects, limiting their applicability.

Purpose of the Study:

• To investigate the transferability of the self-consistent field neural network (SCFNN) model.
• To assess the ability of SCFNN to capture long-range electrostatic phenomena.
• To explore the prediction of nonlinear dielectric response in water.

Main Methods:

• Utilizing a self-consistent field neural network (SCFNN) model.
• Simulating dielectric saturation in water under high electric fields.
• Analyzing nuclear and electronic structure changes during dielectric saturation.

Main Results:

• The SCFNN model successfully predicted dielectric saturation in water.
• The model demonstrated transferability beyond the linear response regime.
• Accurate predictions were achieved without training on high-field data.

Conclusions:

• Neural network potentials, when incorporating relevant physics, can achieve transferability.
• SCFNN models show promise for simulating complex phenomena in polar systems.
• This approach enables genuine predictions for systems beyond their training data.