Dynamically Polarizable Force Fields for Surface Simulations via Multi-output Classification Neural Networks.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Classical Molecular Dynamics (MD) force fields often neglect electronic polarization.
• Accurate simulation of electronic polarization is crucial for modeling solid-liquid interfaces.
• This limitation impacts the prediction of properties in systems like electrochemical energy storage.

Purpose of the Study:

• To introduce a general procedure for incorporating electronic polarization into classical MD force fields using a Neural Network (NN) model.
• To validate the NN/MD model for simulating solid-liquid interfaces, specifically a graphene electrode with an aqueous electrolyte.
• To assess the model's accuracy against Quantum Mechanics/Molecular Dynamics (QM/MD) simulations.

Main Methods:

• Development of a multi-input, multi-output Neural Network (NN) model to treat surface polarization as a classification problem.
• Integration of a custom loss function to enforce physically motivated constraints within the NN.
• Application of the NN/MD framework to a graphene-electrolyte interface system.

Main Results:

• The NN/MD model achieved high accuracy in predicting surface polarization effects.
• The model demonstrated versatility in handling different surface charge states.
• Simulations showed satisfactory agreement with established QM/MD methods.

Conclusions:

• The developed NN/MD approach provides a versatile and accurate method for including electronic polarization in classical simulations.
• This advancement is significant for modeling interfacial phenomena relevant to energy storage technologies.
• The framework offers a computationally efficient alternative to QM/MD for certain applications.