Machine learning analysis of self-assembled colloidal cones.

Area of Science:

• Materials Science
• Computational Science
• Optics and Photonics

Background:

• Automated analysis of microscale colloidal particle self-assembly is crucial for understanding material properties.
• Manual quantification of self-assembled structures from microscopy images is time-consuming and prone to error.

Purpose of the Study:

• To investigate the efficacy of machine learning for accelerating and enhancing the accuracy of quantifying colloidal particle self-assembly.
• To apply deep learning, specifically RetinaNet, to confocal microscopy images of self-assembled colloidal cones.

Main Methods:

• Utilized confocal microscopy to image two-photon lithographed colloidal cones.
• Employed RetinaNet, a convolutional neural network, for identifying self-assembled cone structures.
• Generated synthetic confocal image data using Blender, incorporating z-axis slicing and Gaussian noise, to augment experimental training data.

Main Results:

• A machine learning model trained on a combination of synthetic and experimental data achieved a mean Average Precision (mAP) of approximately 85%.
• The model accurately quantified the degree of self-assembly and the distribution of stack sizes for various cone diameters.
• Performance was optimized by using a hybrid dataset, demonstrating the value of synthetic data generation.

Conclusions:

• Machine learning, particularly RetinaNet with hybrid training data, offers a significant improvement in speed and accuracy for analyzing colloidal self-assembly.
• Synthetic data generation is a viable strategy to enhance machine learning model performance in microscopy image analysis.
• Minor discrepancies highlight areas for future improvement in synthetic data quality and model refinement for diverse particle sizes.