Abstract

The widespread adoption and implementation of two-dimensional (2D) materials are hindered by the challenge of precisely controlling the number of atomic layers during growth. To address this issue, we propose a lightweight model, 2D-TLK, designed for segmenting and identifying the thicknesses and sizes of atomic layer flakes in optical microscopy images. This model utilizes FastViT as the encoder and integrates LRASPP with Knet as the decoder. The 2D-TLK model was trained on a dataset 134 images of molybdenum disulfide (MoS2) flakes with varying in thicknesses, achieve remarkable accuracy of 95.48%, a mean Intersection over Union (mIoU) of 81.23%, and faster inference times, with performance metrics recordingrapid inference speeds of 57.4 FPS (frames per second) on graphical processor unit (GPU) and 1.80 FPS on central processor unit (CPU). Additionally, successful adaptation to WS2 and graphene images confirms its generalizability to different 2D materials. Moreover, we introduce a multi-agent system to enhance interactivity and analytical efficiency. In this system, a Visual Agent collaborates with the 2D-TLK model to identify microscopy images, while a Coder Agent, equipped with a Code Interpreter, processes the recognition results. The system intelligently allocates tasks by dynamically selecting the most suitable agent based on user input, while offering natural language explanations for an efficient and intuitive interaction. This study advances model-driven material characterization and enables AI-assisted scientific discovery by linking computational intelligence with experimental materials science. The code is publicly available at https://github.com/zhouruiliangxian/2D-TLK. © 2025 Elsevier B.V.