Cross-property deep transfer learning framework for enhanced predictive analytics on small materials data.

Area of Science:

• Materials Science
• Computational Materials Science
• Data Science

Background:

• Artificial intelligence (AI) and machine learning (ML) are vital for accelerating materials discovery.
• Deep learning (DL) and transfer learning (TL) show promise but are limited by the need for large datasets, which are often unavailable for many material properties.

Purpose of the Study:

• To develop a cross-property deep transfer learning framework to address the challenge of limited data in materials science.
• To enable the application of DL/TL to material properties with small datasets by leveraging models trained on larger, related datasets.

Main Methods:

• A novel deep transfer learning framework was developed.
• The framework was tested on 39 computational and 2 experimental datasets.
• Models were trained using elemental fractions as input and compared against traditional ML/DL models using physical attributes.

Main Results:

• The transfer learning models, using only elemental fractions, outperformed traditional ML/DL models trained from scratch on 69% of computational datasets and 100% of experimental datasets.
• This suggests transfer learning is highly effective even with limited, simple input features.

Conclusions:

• The proposed deep transfer learning framework effectively tackles the small data challenge in AI/ML for materials science.
• This approach offers a viable solution for building accurate predictive models when large datasets are unavailable, broadening AI/ML applications in the field.