AI electron microscope: 4 servers turn words into experiments

AI electron microscope: 4 servers turn words into experiments Even researchers who arrive with a semester of electron microscopy coursework can need weeks of hands-on training before they run advanced techniques. A 2026 study from Lawrence Berkeley National Laboratory's Molecular Foundry tries a shortcut: an AI assistant that runs your electron microscope when you ask in plain language. The framework is called TEM Agent, and the research appears in npj Computational Materials. It connects a commercial language model to the instrument through the Model Context Protocol, a standard that links assistants with specific tools. The team had already written code to automate individual tasks, and this work made that toolkit available to the AI. You don't hand it the keys. The assistant can only choose from approved actions, such as adjusting magnification, focusing or tilting a sample, and combine them into longer workflows. That means no fresh code for every new experiment. Since tomography creates a 3D image from images acquired with various angles of the sample, all of the steps need to occur in the proper order. In the first attempt of the assistant, the change was made in the scanning direction, not in the sample orientation. An explanation of the differences was provided to the researchers, and the process could continue. Additionally, they have tried a more challenging approach for sample reconstruction using patterns of scattered electrons. It is important. Simple control requires an expert nonetheless. Links to the NERSC supercomputing centre and the Crucible database extend the process beyond imagery capture. It is more efficient and allows the extraction of previous experiments' parameters without searching for independent documentation. Berkeley Lab claims that it reduced the turnaround time from hours to minutes. It does not mean that the outcome demonstrates an ability to interpret each instruction perfectly and makes experts redundant. It means that there is a restricted setup and one identified fix.