Deep Learning-Based Prediction of Material Properties Using Chemical Compositions and Diffraction Patterns as...

Area of Science:

• Materials Science
• Computational Materials Science
• Machine Learning

Background:

• Predicting material properties is crucial for discovering new materials.
• Current deep learning models often require complex, not directly accessible input data.
• There is a need for models that utilize readily available experimental data.

Purpose of the Study:

• To develop a deep learning model for predicting material properties using accessible experimental inputs.
• To integrate heterogeneous data sources (chemical composition and diffraction data) within a single model.
• To demonstrate the practical utility of deep learning in materials science by avoiding inaccessible parameters.

Main Methods:

• Developed a deep learning model capable of processing heterogeneous inputs: chemical composition and diffraction data.
• Introduced a novel chemical composition vector using element embedding and a normalized composition matrix.
• Trained the model on 1524 binary samples from the Materials Project database.

Main Results:

• The model accurately predicts formation energies (MAE: 0.29 eV/atom) and band gaps (MAE: 0.66 eV).
• Analysis indicated that chemical composition has a greater influence on predicted properties than crystal structure.
• Successfully avoided the use of non-directly accessible inputs like atomic coordinates.

Conclusions:

• The developed deep learning model effectively predicts material properties from routine experimental data.
• The model's reliance on accessible inputs significantly enhances its practical applicability in materials research.
• Chemical composition is identified as a dominant factor influencing material properties compared to crystal structure.