Machine Learning-Enabled Image Classification for Automated Electron Microscopy.

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Traditional materials discovery relies on intuition, lacking systematic design.
• Advancements in big data and computational power enable AI and ML for accelerated materials discovery.
• Combinatorial megalibraries necessitate automated characterization tools for nanoparticle analysis.

Purpose of the Study:

• To develop a machine learning (ML) model for real-time binary classification of nanoparticle images.
• To minimize false positives in nanoparticle classification, reducing downstream processing errors.
• To address computational challenges in ML model development for materials discovery.

Main Methods:

• Development of a specialized ML model for binary classification of grayscale high-angle annular dark-field images.
• Implementation of strategies to manage memory constraints and optimize training time.
• Utilization of Neural Architecture Search tools for model optimization.

Main Results:

• The ML model achieved over 95% precision.
• The model demonstrated a weighted F-score exceeding 90% on test data.
• The developed model effectively classifies nanoparticles in real-time.

Conclusions:

• AI and ML significantly accelerate the discovery of novel materials.
• The developed ML model represents a significant advancement in applying AI to materials discovery.
• The model's high precision and efficacy address critical needs in automated materials characterization.