Neural Network Force Fields for Metal Growth Based on Energy Decompositions

Area of Science:

• Computational Materials Science
• Machine Learning in Physics
• Ab Initio Simulations

Background:

• Accurate simulation of metal growth is computationally expensive using traditional methods like ab initio density functional theory (DFT).
• Existing machine learning (ML) methods may not provide sufficient detail or require extensive data for accurate materials modeling.
• Developing efficient and accurate simulation techniques is crucial for understanding and predicting material properties and behavior.

Purpose of the Study:

• To propose a novel ML-based method for simulating metal growth that combines DFT accuracy with reduced computational cost.
• To leverage atomic energy decomposition from DFT calculations for enhanced information yield in ML models.
• To validate the developed ML approach by comparing its simulation results with DFT and experimental observations.

Main Methods:

• Utilizing atomic energy decomposition derived from DFT calculations as the foundation for the ML model.
• Developing a neural network potential (NNP) trained on a limited dataset (1000 DFT molecular dynamics images) of amorphous sodium.
• Comparing simulation dynamics and structural properties generated by DFT and the trained NNP.
• Conducting metal growth experiments from liquid to solid states in varying system sizes to test NNP simulation capabilities.

Main Results:

• The proposed energy decomposition ML approach yields significantly more information compared to other ML methods using the same DFT calculations.
• An accurate ML model was trained for the amorphous sodium system using only 1000 DFT molecular dynamics images.
• Simulations using DFT and the developed NNP produced comparable structural properties and dynamics.
• The NNP successfully simulated real metal growth processes from liquid to solid states in both small and large systems.

Conclusions:

• The developed ML method based on DFT energy decomposition offers a computationally efficient alternative to traditional DFT for materials simulations.
• The NNP approach accurately captures the structural properties and dynamics of metal systems, including the complex process of metal growth.
• This work demonstrates the potential of ML, specifically NNP, to simulate realistic materials growth phenomena, paving the way for accelerated materials discovery and design.