The way we build, operate, and think about microscopes is changing. After more than a decade in the field, both as a microscopist and now as a manager of microscopy infrastructure, I am seeing two trends converge more decisively than ever. AI is arriving across the entire stack. Automation is no longer an industrial luxury but a practical research tool. The two reinforce each other.
The convergence is happening from every direction
The interesting part is that this is not coming from one place. Commercial vendors are shipping production tools: Nikon’s NSPARC and JOBS pipelines, Zeiss arivis analysis-on-instrument workflows, Abberior’s MINFLUX automation, Thermo Fisher’s autonomous electron microscopy stacks. National infrastructure programmes are putting serious budgets behind integrated AI-driven materials discovery: DIADEM in France (€85M), CAPeX in Denmark (€45M), INTERSECT at DOE in the US, the Korean 500-lab target, and the DP-Technology / AISI-Beijing stack in China (Bohrium platform, Uni-Lab-OS orchestration, DeepModeling models including DPA-3 and OpenLAM). Universities are publishing reference architectures for self-driving labs at a steady cadence. And a growing community of highly motivated individuals is building open-source frameworks (HELAO, ChemOS, AlabOS, IvoryOS, AtomAI, gpCAM, BlueSky / Ophyd) that turn yesterday’s expensive autonomy into something a postdoc with a Python kernel can run today.
Layered on top of all this is the recent jump in agentic AI. Opus 4.6 and 4.7, Claude Code, Codex: these are not the same tools we had even six months ago. They plan, they write the glue code, they read instrument metadata, they propose the next experiment, and they pause and ask when they are unsure. The work has not vanished. It has shifted, from typing the macro to designing the experiment.
A new contract between instrument makers, microscopists, and AI
What I think we are entering is a new contract between the people who design instruments, the people who use them, and the people who build the AI that sits between. Microscopists and instrument makers used to talk past each other across long product cycles. Software engineers arrived late, if at all, to bolt analysis onto a finished system. With agentic AI in the loop, this division stops making sense. The most useful microscopes of the next decade will be co-designed: the optical engineer, the imaging scientist, the data engineer, and the AI agent, sitting at the same table from the start. A few groups are already working this way. Most are not.
This is the moment to be explicit about two things.
1. Collaborate openly
If you have an idea for an experiment that needs three modalities your group does not own, say so. If you have a control framework that you think could replace proprietary scripting on the next confocal, share it. If you suspect agentic AI could rewrite how a facility is operated, propose a pilot. The invisible barriers (who owns the instrument, whose grant the postdoc sits on, which group wrote the paper first) matter less than they used to, and they will matter even less in five years. We waste real time policing them. The interesting opportunities live at intersections, and intersections need explicit invitation.
2. Design AI that helps, not AI that displaces
Agentic systems will absorb a lot of the operational burden of running modern microscopy facilities: booking, training, drift correction, parameter tuning, data triage, even drafting first-pass interpretations. This is good. It frees scientists to do the parts that need human judgement: framing the question, designing the experiment, deciding whether the answer makes physical sense. If we let AI do those parts unchecked, we have built a worse version of science. If we let AI handle the operational parts well, we have built infrastructure that returns our most scarce resource, which is attention. The difference between those two outcomes is being decided right now, by who shows up at the design table.
From AMF, with an open invitation
I am writing this from the Advanced Microscopy Facility at TU/e. We are working on our piece of this: integrating multimodal characterisation into self-driving lab loops, making instrument access transparent, building partnerships with vendors so their systems speak the same data language as our home-built ones, exploring how agentic AI changes day-to-day operation. We do not have all the answers. We do not need to. What we do need is people willing to talk across institutional boundaries.
If you have a project, an instrument, an idea, a frustration, or a half-formed proposal that touches AI, automation, and microscopy, get in touch. Email superresolution@tue.nl, or write to me directly at y.wang8@tue.nl. Tell us what you want to build. We will tell you what we are building. Some of it will line up. Some of it will not. Either way, that beats waiting for the perfect moment.
The next decade of microscopy infrastructure will be built by collaboration, not by individual brilliance. The technology is finally ready. The question is whether we, as a community, are.
— Yuyang Wang, Head of AMF, Strategy Lead at ICMS / TU/e