Enabling deeper learning on big data for materials informatics applications.

Area of Science:

• Materials Science
• Machine Learning
• Deep Learning

Background:

• Machine learning (ML) is increasingly used in materials science for data-driven property prediction.
• Traditional ML is common, but deep learning (DL) is limited by data scarcity and the vanishing gradient problem in deep networks.
• Despite challenges, the potential of DL and growing datasets make deeper networks attractive.

Purpose of the Study:

• To develop a deep learning framework enabling deeper neural networks for materials science property prediction.
• To address the vanishing gradient problem in deep learning models when large materials datasets are available.
• To improve model accuracy beyond traditional ML and plain deep networks.

Main Methods:

• A general deep learning framework, Individual Residual learning (IRNet), was developed.
• IRNet utilizes very deep neural networks capable of processing vector-based materials representations.
• The framework is designed to mitigate the vanishing gradient problem.

Main Results:

• IRNet successfully alleviates the vanishing gradient problem, enabling the use of deeper neural networks.
• The proposed IRNet models achieve significantly better accuracy (up to 47%) compared to plain deep networks.
• IRNet outperforms traditional ML techniques for property prediction with big data.

Conclusions:

• Individual Residual learning (IRNet) provides an effective approach for deep learning in materials science.
• The framework facilitates the development of highly accurate property prediction models using large datasets.
• IRNet represents a significant advancement in applying deep learning to materials discovery and design.