Berkeley Lab’s AI assistant takes natural-language instructions at the electron microscope

Researchers at Lawrence Berkeley National Laboratory have built an AI assistant that can operate a transmission electron microscope from natural-language instructions. TEM Agent connects the instrument to a supercomputer and an experiment-data platform, helping researchers get from request to results without needing to hand-program every sequence of steps.

Watch Desk analysis

What happened

The team at Berkeley Lab’s Molecular Foundry connected a large language model to existing microscope automations through the Model Context Protocol (MCP). Rather than granting open-ended control, MCP exposes a fixed set of pre-written actions. TEM Agent can combine those actions to adjust magnification, tilt a sample and refocus the beam, among other tasks.

The team tested the system on established techniques including electron tomography, which builds a three-dimensional view from tilted images, and on ptychography, a more experimental imaging method. The report says TEM Agent also connects the microscope to the NERSC supercomputer and the Crucible data platform. Researchers can process raw data in seconds rather than hours and ask for settings used in earlier experiments instead of digging through files or lab notebooks. The work is described in a study published in npj Computational Materials.

Why it matters

Advanced microscopes can be difficult to use, and their detectors can produce around 700 gigabytes of data in 15 seconds. Letting researchers describe a task in ordinary language could make specialist equipment and techniques more accessible, while the links to computing and experiment records reduce some of the manual hand-offs around a session.

The safety-relevant detail is in the design: the model gets a set of specific, pre-written actions, not unrestricted control of the microscope. That does not make every experiment foolproof. The researchers encountered an early error in which the system confused tilting a sample with rotating the scan direction, then corrected it with an explanation.

Our read

This is a useful example of an AI agent doing more than chatting about science: it can coordinate real equipment through a bounded toolkit. The interesting achievement is not that plain English has replaced expertise, but that existing automation can be made easier to use and combine. The practical test now is how reliably that works across experiments and instruments beyond the team’s initial setup.

What to watch

• Whether TEM Agent handles a wider range of experiments reliably, especially when a run goes off the expected path.
• Whether the MCP and Crucible approach extends to other instruments and research centres.
• How researchers keep human oversight meaningful when an assistant chains microscope actions together.

Discussion spark: If an AI assistant can operate complex scientific equipment through a fixed set of approved actions, should that make advanced experiments accessible to more researchers, or keep operation firmly in specialist hands?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.