Machine learning approaches for elastic localization linkages in high-contrast composite materials.

Area of Science:

• Materials Science
• Data Science
• Computational Modeling
• Machine Learning

Background:

• Advanced data science and informatics offer solutions for computational challenges in multiscale materials science modeling and simulation.
• Mining microstructure-property relationships using machine learning (ML) and data mining (DM) presents opportunities for rapid material design.

Purpose of the Study:

• To explore and present computationally efficient methods for predicting microscale elastic strain fields in 3D voxel-based microstructure volume elements (MVEs).
• To demonstrate improvements in prediction accuracy and efficiency through advanced ML/DM techniques.

Main Methods:

• Exploration of ML/DM concepts including feature extraction, feature ranking and selection, and regression modeling.
• Development and application of feature descriptors representing voxel neighborhood characteristics.
• Utilized a reduced set of top-ranked descriptors and an ensemble-based regression technique.

Main Results:

• Demonstrated gradual improvements in prediction efficiency and accuracy through an escalated approach.
• Identified key feature descriptors crucial for accurate strain field prediction.
• Validated the effectiveness of the ensemble-based regression technique for MVE analysis.

Conclusions:

• Advanced ML/DM approaches provide computationally efficient pathways for predicting microscale elastic strain fields.
• The proposed methods, including feature selection and ensemble regression, significantly enhance material design capabilities.
• This work highlights the potential of data-driven strategies in accelerating materials science research and development.