SEMPro: A Data-Driven Pipeline To Learn Structure-Property Insights from Scanning Electron Microscopy Images.

Area of Science:

• Materials Science
• Biomaterials Engineering
• Data Science

Background:

• Hydrogel microstructure analysis via scanning electron microscopy (SEM) is vital for understanding material properties.
• Current methods rely on subjective interpretation and limited datasets, hindering comprehensive structure-property relationship analysis.

Purpose of the Study:

• To develop a data-driven solution, SEMPro, for automated analysis of hydrogel SEM images.
• To establish a pipeline for compiling and analyzing hydrogel structure-property relationships using deep learning (DL).

Main Methods:

• Web-scraping techniques to compile a large dataset of hydrogel SEM images.
• Deep learning models, including transfer learning and activation mapping, for image analysis and property prediction.
• Explainable AI (XAI) methods to validate model predictions and interpret feature relevance.

Main Results:

• SEMPro accurately predicts the elastic modulus of hydrogels from SEM images within the same order of magnitude.
• The model successfully extracts modulus-relevant microstructural features, visualized through activation mapping.
• Explainable AI confirmed the model's predictive validity and feature importance.

Conclusions:

• SEMPro offers a closed-loop system for hydrogel data collection and analysis, enabling high-dimensional insights.
• This approach can significantly advance the understanding and design of hydrogels and soft materials.
• The integration of DL and XAI provides a powerful tool for materials science research.