Machine learning material properties from the periodic table using convolutional neural networks.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Convolutional neural networks (CNNs) excel at feature extraction in image recognition.
• The periodic table encodes rich chemical information crucial for predicting material properties.

Purpose of the Study:

• To investigate CNNs' ability to learn chemical information directly from the periodic table structure.
• To predict material properties like lattice parameters and formation enthalpy.
• To identify potentially stable novel compounds using transfer learning.

Main Methods:

• A multi-task CNN was trained using full-Heusler compounds from the Open Quantum Materials Database (OQMD).
• The periodic table served as the input representation.
• Transfer learning was applied by fine-tuning the CNN on a second dataset from the Inorganic Crystal Structure Database (ICSD).

Main Results:

• CNN predictions achieved accuracy within DFT (density functional theory) precision for lattice parameters and formation enthalpy.
• The CNN learned the inherent chemical information from the periodic table's 2D structure.
• Transfer learning successfully predicted the stability of full-Heusler compounds.
• Tungsten-containing compounds were identified as potentially stable and rarely reported.

Conclusions:

• CNNs can effectively extract and utilize chemical information encoded in the periodic table.
• This data-driven approach offers a powerful tool for materials discovery and property prediction.
• The study highlights the potential of transfer learning for identifying novel, stable materials.