I was asked to be on a panel recently to discuss self-driving laboratories with
@ianfoster and Jie Xu (UC/ANL). As part of this exercise I distilled down the following thoughts:
Bad scenarios:
We might get more, but not necessarily better: not everything is amenable to the SDL paradigm (at least today). A risk to be avoided is that work that can’t be automated languishes because it is too hard and not hip to support anymore. This is a bit like AI sucking the air out of the room for other useful investment today.
Not everything needs to be automated: SDL talks will sometimes feature an image of a laboratory with robots and no humans that triggers some people. This is a sticky issue. I see no virtue in pursuing automation just for the sake of getting people out of the lab, but there are real costs associated with normal human sloppiness. Removing people from the interpretation, communication, planning, and other higher level processes associated with research, seems less urgent.
We destroy the entry level ladder to research in the physical sciences and engineering and we don’t replace it.
Good scenarios:
In applicable areas, SDLs will bring more reproducible, transparent, and accelerated discovery cycles.
The distinction between simulations and experiments will collapse into a much more unified set of tools for testing hypotheses.
Reduced activation energy for testing ideas. If you talk to a seasoned researcher they probably have many untested ideas. Most are probably untested not because they are particularly hard but because no one was available at the time or they weren’t funded for it, or a million other inessential things.
Healthy reconsideration of how we allocate prestige in science and engineering. SDLs are part of the myriad forces arrayed against conventional publishing and research organization.