A Universal 3D Voxel Descriptor for Solid-State Material Informatics with Deep Convolutional Neural Networks.

Area of Science:

• Materials Science
• Computational Chemistry
• Data Science

Background:

• Material informatics (MI) accelerates material discovery using machine learning, but existing descriptors struggle with 3D solid-state data.
• Insufficient representation of 3D field quantities like electron distributions limits the practical success of solid-state MI.

Purpose of the Study:

• To develop a generic 3D voxel descriptor for solid-state material informatics.
• To enable the effective use of convolutional neural networks (CNNs) in solid-state MI by improving feature representation.

Main Methods:

• A simple, generic 3D voxel descriptor was developed to encode field quantities.
• The descriptor was tested using electron distribution data from 680 oxide materials.
• Regression analysis was performed to predict Hartree energies.

Main Results:

• The proposed 3D voxel descriptor outperformed existing methods in predicting Hartree energies.
• The descriptor effectively captures the long-wavelength distribution of valence electrons.
• The scheme demonstrates the potential to predict various functionals of field quantities with sufficient data.

Conclusions:

• The new 3D voxel descriptor significantly advances solid-state material informatics.
• This approach facilitates the integration of advanced machine learning techniques, including supervised, semi-supervised, and reinforcement learning, into solid-state MI.
• It paves the way for more efficient and accurate material discovery processes.