Machine Learning Enhanced Optical Characterization of Hexagonal Boron Nitride Thickness on 300-nm Oxide Substrate

This study integrates optical techniques with machine learning to achieve precise characterization of hexagonal boron nitride (hBN) thickness on 300-nm silicon oxide substrate. Conventional methods for interpreting large volumes of microscopy data are labor intensive, time consuming, and faulted by human subjectivity. The aim of this work is to mitigate these challenges by employing a regional convolutional neural network (R-CNN) to perform image segmentation and object detection on optical microscopy images, enabling real-time processing capabilities for thickness identification. The Mask R-CNN architecture featuring the ResNet101 backbone network, region proposal network, classification and mask subnets with bounding box regression generates feature maps to identify graphical features including shapes, flake sizes, contrast, and color. Initially, a robust theoretical mapping of hBN flake colors in the standard red, green, blue (sRGB) color space was developed, accounting for the oxide thickness. Theoretical RGB values corresponding to specific hBN thickness align remarkably well with those extracted from optical images, validating the use of colors as a reliable standard for hBN thickness identification. Our trained R-CNN predicts thickness of hBN flakes based on their mapped expected color variations with remarkable accuracy. This material characterization approach accelerates initial characterization of hBN while also holding potential for enhancing production efficiency across various 2D materials in nanoscience research.