MatImageAgent: A Large Language Multimodal Agent for Materials Characterization Images

ABSTRACT Our study introduces MatImageAgent, an agent‐based system that employs large language models (LLMs) to completely automate the image analysis workflow in materials science. MatImageAgent was developed to process images from scanning electron microscopy (SEM), X‐ray computed tomography (XCT), and atomic force microscopy (AFM) using properly engineered task descriptions (TDs) to execute an integrated process from image recognition and quantitative analysis to automated report generation. Its performance was evaluated using three primary tasks: detecting Li‐striped voids and estimating heights in SEM imaging (TASK 1), extracting greyscale values and analyzing lithium deposition in XCT data (TASK 2), and separating different phases in polymer films using AFM scans (TASK 3). MatImageAgent was coupled with three popular LLMs (GPT‐4.1, Claude 3.7, and DeepSeek‐R1). It exhibited high stability up to 10 independent cycles, yielding a mean absolute error of 0.11 μm for SEM measurements and an accuracy of 0.82 for AFM phase marking. Our AI agent improves analytical productivity by minimizing the requirement of advanced programming skills.