Image processing pipeline for AI-driven nanoparticle megalibrary characterization.

Area of Science:

• Materials Science
• Nanotechnology
• Artificial Intelligence

Background:

• Megalibraries enable the production of millions of unique nanoparticles on a chip.
• Vast datasets generated by megalibraries require automated analysis tools.
• Previous work developed a machine learning model for nanoparticle image quality selection.

Purpose of the Study:

• To develop an automated tool for analyzing nanoparticle images from megalibraries.
• To improve the performance and robustness of machine learning models for nanoparticle characterization.
• To reduce the time and cost associated with analyzing large nanoparticle datasets.

Main Methods:

• Implemented a custom image processing pipeline before training machine learning models.
• Utilized the pipeline to clean and enhance raw nanoparticle images.
• Trained binary classification models using processed images, including lower-resolution data.

Main Results:

• The image processing pipeline significantly improved model performance, with an 18.2% increase in recall and 13.1% increase in accuracy.
• The best model achieved 95.9% precision and a 95.1% weighted F-score on an unseen test set.
• Model training time was reduced from hours to under a minute, and performance improved at lower resolutions.

Conclusions:

• The custom image processing pipeline enhances AI model performance for nanoparticle characterization.
• This approach enables faster, more accurate analysis of large nanoparticle datasets, accelerating materials discovery.
• The pipeline offers cost savings and improved robustness to image variations.