Learning properties of ordered and disordered materials from multi-fidelity data.

Area of Science:

• Materials Science
• Computational Materials Science
• Machine Learning

Background:

• Predicting material properties from atomic structure is crucial.
• Machine learning (ML) offers rapid predictions but is hindered by limited high-fidelity data.

Purpose of the Study:

• To develop a universal approach for accurate materials property prediction with small datasets.
• To enhance the utility of ML in materials science by addressing data scarcity.

Main Methods:

• Developed multi-fidelity graph networks (MFGNs).
• Incorporated low-fidelity Perdew-Burke-Ernzerhof (PBE) band gaps into graph networks.
• Utilized learned elemental embeddings for modeling material disorder.

Main Results:

• MFGNs achieve accurate predictions with small data sizes.
• Inclusion of low-fidelity PBE band gaps reduced mean absolute errors by 22-45% for experimental band gap predictions.
• Learned elemental embeddings effectively model disorder in materials.

Conclusions:

• Multi-fidelity graph networks provide a powerful solution for data-scarce materials property prediction.
• This method significantly improves the accuracy of computational materials predictions.
• The approach addresses a key challenge in computational materials science, particularly for disordered materials.