From Instruments4AI to AI4Instruments — an invitation to discuss and build

[Cherish0925]

I started AI4Instruments after working on AI systems that carry out research on real experimental apparatus. A question kept coming back: how much of an experiment is shaped by the physical capabilities of an instrument, and how much by the small part of those capabilities its software lets us use?

Getting an AI system to operate existing equipment already takes substantial work: drivers, manuals, interfaces, calibration, and coordination between devices. I call that starting point Instruments4AI. AI4Instruments asks what becomes possible when AI also participates in designing and developing the instruments themselves, from the initial scientific question and physical design through manufacture, circuits, controllers, firmware, software, experimental use, and subsequent redesign.

The idea is AI design and implementation + AI development and use. An experiment can reveal a missing capability; developing that capability can make a different experiment possible. Sustained autonomous research should be able to work through both sides of this relationship.

One part of the proposal concerns operations. Basic Operations expose the elementary controls and observations available at a defined hardware level. Standard Operations express reusable experimental procedures. Workflow Operations compose them across devices, laboratories, and time. Existing procedures should remain open to inspection, modification, and replacement. A new research question may require a different composition, direct access to a controller, or a change to the apparatus itself.

I want the shared descriptions to preserve what is distinctive about each instrument: its physical behaviour, timing, measurement conditions, and degrees of freedom. The scope extends across scientific and engineering instruments and across different kinds of AI. Experimental records also matter here. Connecting designs, actions, observations, and research decisions could help laboratories build on one another's work and give AI systems useful experience from the physical world.

The repository currently contains the paper, design notes, and a small executable optical simulation. The simulation illustrates the operation model; it is one starting point for examining a much broader proposal.

Read the paper · PDF · Operations · Executable example

I would like to develop this with people who build instruments, write their software, and use them for research. A few questions I would particularly value your experience on:

• Where has an interface prevented an experiment that the hardware could otherwise support?
• What should be shared between instruments, and what needs to remain specific to their physics and construction?
• What experiment would make a convincing test of this approach, including the possibility that parts of it need to be changed?

A concrete example, a counterexample, a design sketch, or a reference I have missed would all help. You can reply here or open a separate discussion for an instrument or research question. English and Chinese are both welcome.

Shuxing Yang