Matini-Net: Versatile Material Informatics Research Framework for Feature Engineering and Deep Neural Network Design.

Area of Science:

• Materials Science
• Computer Science
• Data Science

Background:

• Materials informatics leverages data science and machine learning to accelerate materials discovery.
• Deep learning offers powerful tools for materials informatics but requires specialized expertise.
• Automating feature engineering and model design is crucial for broader adoption.

Purpose of the Study:

• Introduce Matini-Net, a versatile framework for automated deep learning model design and feature engineering in materials informatics.
• Enable researchers with limited deep learning experience to effectively apply machine learning to materials research.
• Enhance model interpretability through automated feature importance analysis.

Main Methods:

• Developed Matini-Net, a flexible framework supporting feature-based, graph-based, and hybrid deep learning models.
• Designed single- and multimodal model architectures.
• Validated performance on the MatBench benchmarking dataset across five material properties using regression architectures.

Main Results:

• Achieved R-squared (R2) values greater than 0.84 on five material property datasets.
• Demonstrated the framework's flexibility in designing various regression architectures.
• Successfully applied automated feature engineering, hyperparameter tuning, and network construction.

Conclusions:

• Matini-Net significantly accelerates materials discovery by simplifying deep learning application.
• The framework enhances model interpretability, aiding in understanding material-property relationships.
• Matini-Net promotes wider and more effective use of machine and deep learning in materials research.