Computer Vision, Machine Learning + Materials Engineering

This workflow demonstrates a computer vision and convolutional neural network (CNN)-based framework for classifying assembled upconversion nanoparticle (UCNP) layers from scanning electron microscopy (SEM) images of microsphere surfaces. Through synthesis control, the ratio between bottom (B) and second (S) layered UCNPs was systematically tuned to engineer different assembly structures. SEM imaging data were collected from these self-assembled nanoparticle systems and manually labeled to generate training datasets for AI-assisted analysis. The image analysis pipeline included preprocessing, particle detection and segmentation, feature extraction, and CNN training for automated layer classification. The trained model successfully identified and mapped B and S particles across new SEM datasets obtained at multiple magnifications (50k×, 80k×, and 100k×), demonstrating robust generalization and enabling large-scale quantitative structural analysis of nanoparticle assemblies through computer vision and deep learning. 

Project 2. Reinforcement Learning-Driven Classficiation of Gold Nanoplates without Training Data

This schematic demonstrates an RL-based framework for classifying triangle-like (blue) and hexagon-like (green) Au nanoplates from optical microscopy images without any pre-labeled training dataset. The agent learns shape-classification policies from scratch, receiving a reward signal based on agreement between its predicted label and expert-corrected ground truth, progressively refining its detection and classification accuracy through iterative feedback. As training progresses (left to right), detectability and correctness rise from 45%/15% to 89%/75%. Red-boxed insets mark misclassified cases from earlier episodes (e.g., a triangle mislabeled as "hexagon-like"), illustrating how reward-driven refinement corrects these errors over successive iterations.