Abstract
Ptychography is an advanced computational imaging technique in X-ray andelectron microscopy. It has been widely adopted across scientific researchfields, including physics, chemistry, biology, and materials science, as wellas in industrial applications such as semiconductor characterization. Inpractice, obtaining high-quality ptychographic images requires simultaneousoptimization of numerous experimental and algorithmic parameters.Traditionally, parameter selection often relies on trial and error, leading tolow-throughput workflows and potential human bias. In this work, we develop the"Ptychographic Experiment and Analysis Robot" (PEAR), a framework thatleverages large language models (LLMs) to automate data analysis inptychography. To ensure high robustness and accuracy, PEAR employs multiple LLMagents for tasks including knowledge retrieval, code generation, parameterrecommendation, and image reasoning. Our study demonstrates that PEAR'smulti-agent design significantly improves the workflow success rate, even withsmaller open-weight models such as LLaMA 3.1 8B. PEAR also supports variousautomation levels and is designed to work with customized local knowledgebases, ensuring flexibility and adaptability across different researchenvironments.