Multimodal machine learning for integrating heterogeneous analytical systems.

Area of Science:

• Materials Science
• Machine Learning
• Data Analysis

Background:

• Understanding complex material properties requires integrating multi-scale measurements.
• Carbon nanotube (CNT) films exhibit properties sensitive to microstructural variations.
• Existing characterization methods may not fully capture intricate structure-property relationships.

Purpose of the Study:

• To develop an interpretable multimodal machine learning framework for end-to-end characterization of complex materials.
• To unify heterogeneous analytical systems for comprehensive material analysis.
• To demonstrate the framework's efficacy on carbon nanotube films.

Main Methods:

• Extracted quantitative morphology descriptors from SEM images (binarization, skeletonization, network analysis).
• Fused SEM features with Raman indicators, specific surface area (gas adsorption), and electrical surface resistivity.
• Employed multi-dimensional visualization (radar plots, UMAP) for data clustering.
• Utilized nonlinear regression models (XGBoost) for predictive accuracy and feature importance analysis.

Main Results:

• Multimodal data visualization revealed clear clustering of CNT films based on crystallinity and entanglements.
• XGBoost models achieved high predictive accuracy, outperforming other nonlinear approaches.
• Feature importance analysis provided physically meaningful interpretations of structure-property correlations.
• Identified key descriptors governing surface resistivity and specific surface area.

Conclusions:

• The proposed multimodal machine learning framework enables data-driven and explainable characterization of complex materials.
• This approach effectively integrates diverse analytical data for deeper material understanding.
• Demonstrated a powerful new methodology for advancing materials science research and development.