Segment Any Flakes— A Quantum Crystal-Based Tokenization Learning Approach to Quantum 2D Material Understanding

Detecting two-dimensional (2D) materials within silicon chips poses a significant challenge in the realm of quantum machines due to the complexities of data collection. Recently, artificial intelligence (AI) and deep learning have demonstrated remarkable success across various fields, and research into 2D quantum materials stands to benefit from these advancements. The Segment Anything Model (SAM) is a notable development in the expansion of segmentation models, offering powerful zero-shot capabilities and flexible prompting. However, despite being trained on 1.1 billion masks, SAM’s mask prediction quality is often insufficient, particularly when dealing with objects featuring intricate structures, such as quantum flakes. This work addresses the limitations of SAM in detecting 2D quantum flakes, particularly its tendency to over-segment flakes due to variations in thickness. This issue arises because SAM treats different regions of a single flake as separate segments, leading to reduced mean average precision (mAP) by generating multiple predictions for a single flake. To address this challenge, we introduce a novel approach that contains three key contributions: (1). A high-quality dataset specifically tailored for quantum 2D flake detection, with detailed annotations based on individual boundaries. (2) A self-supervised learning (SSL) algorithm designed to better capture the intrinsic properties of quantum flakes, enhancing the model's representation and understanding of these materials. (3) The integration of our dataset and SSL approach into SAM resulted in significant performance improvements and a marked reduction in false positives during flake segmentation. Our work lays the foundation for more accurate and reliable segmentation of 2D materials, advancing quantum research.