A lightweight model and agent framework for fast and accurate surface defect detection in MoS2 films

Two-dimensional (2D) materials, particularly transition metal dichalcogenides (TMDs), have shown great potential in optoelectronics and energy storage due to their unique properties. However, defects such as impurities and voids can significantly impact their electronic, optical, and magnetic characteristics. Traditional defect detection methods, such as scanning transmission electron microscopy (STEM), are time-consuming and limited in scalability. This study introduces DeepMGD, a novel deep learning model designed for efficient and accurate defect detection in molybdenum disulfide (MoS${}_2$) films fabricated via chemical vapor deposition (CVD) and imaged using optical microscopy. DeepMGD utilizes MobileNetV4 as its backbone, Gold-YOLO as the neck, and a decoupled head. The architecture attains a competitive mAP@50 of 0.894 under challenging illumination conditions while maintaining a lean parameter count of 5.129 million, achieving 122.2 FPS on a GPU and 3.0 FPS on a CPU. Additionally, we present the DeepMGD Agent, an intelligent system framework that integrates the DeepMGD model with a large language model (LLM) and a Python interpreter. This framework automates defect detection and analysis, offering an intuitive workflow for users. The system processes microscopic images and user commands to detect defects and generate natural language explanations, enabling seamless defect detection and quantitative analysis. This work provides a reliable and efficient approach for analyzing 2D materials, with potential applications for other similar materials in the future. The code is released at https://github.com/zhouruiliangxian/DeepMGD.