Voxelized Atomic Structure Potentials: Predicting Atomic Forces with the Accuracy of Quantum Mechanics Using...

Area of Science:

• Materials Science
• Computational Physics
• Machine Learning

Background:

• Developing accurate interatomic potentials is crucial for simulating materials behavior.
• Traditional methods often struggle with complex atomic structures and multi-body interactions.

Purpose of the Study:

• Introduce the voxelized atomic structure (VASt) potentials as a novel machine learning framework.
• Enable high-fidelity representation of complex atomic environments for accurate force prediction.

Main Methods:

• Utilize a voxelized representation of atomic structures as input for convolutional neural networks (CNNs).
• Employ CNNs to implicitly learn low-dimensional features correlating atomic neighborhoods to net atomic forces.
• Automate feature engineering for scalability and efficiency.

Main Results:

• Demonstrate high-fidelity modeling of spatial relationships and multi-body interactions in atomic systems.
• Achieve highly accurate force predictions for two phases of silicon carbide.
• Accurately predict the thermal conductivity of silicon under isotropic strain.

Conclusions:

• VASt potentials offer an efficient and scalable machine learning framework for interatomic potential development.
• The method shows remarkable fidelity in capturing the physics governing atomic systems.
• VASt potentials are effective for predicting material properties like forces and thermal conductivity.