Before Rainbow's robots synthesized a single nanocrystal, the researchers spent months on a question that sounds like housekeeping and turns out to be the whole job: what should the machine be permitted to try?
Rainbow is a self-driving materials laboratory. Four robots, linked to optimization algorithms, prepare chemical precursors, run reactions, and characterize the results on 96-well plates, finishing each cycle of 24 experimental conditions in under 150 minutes. Across its campaign it ran 3,648 synthesis experiments without human hands in the loop. But the harder work came before any of that.
The team chose a library of six organic acids and ruled out polar solvents because of the ionic character of the nanocrystals they wanted. They translated "make a better nanocrystal" into three measurable properties: a specified peak emission energy, maximum proxy photoluminescence quantum yield, minimum emission linewidth. Candidates predicted to fall outside ±2% of the target emission were scored zero and could never rank high enough to advance. Each of those decisions compresses a chemical judgment into a rule the optimizer can enforce, and each requires knowing enough chemistry to say what is being excluded and why the exclusion is safe.
Then came budget allocation, which has no clean answer. More initial experiments give the model a broader view of the chemical landscape but leave fewer cycles for targeted optimization. Fewer initial experiments preserve optimization cycles but risk bias from sparse data. The team tested several allocations in simulation, settled on a split, then added an extra initialization cycle after roughly 20% of early reactions produced nothing emissive at all. They also built a classifier to recognize the non-emissive results, since a model trained only on successful syntheses will happily steer the search into dead regions of chemical space. That classifier raised the emissive-synthesis rate by 12.5%.
A protein-engineering platform called SAMPLE at the University of Wisconsin–Madison shows the same work in a different material. There the specification problem was constructing the set of proteins the machine was allowed to invent: 1,352 valid sequences assembled from 34 DNA fragments, with recombination breakpoints chosen using structural information. The fragments were designed for broad diversity rather than specifically to improve thermostability, the property under study, which means the researchers were encoding a bet about where useful variation might be hiding. SAMPLE's algorithms searched under 2% of that landscape and found enzymes at least 12°C more stable than the sequences they started from.
Continuous operation that could in principle have finished in two weeks took close to six months, broken up by robotic malfunctions, reagent restocking, and a shipping delay of two and a half months. Field reviews suggest this is ordinary: nearly every published self-driving lab still needs human maintenance between campaigns and heavy intervention when instruments fail.
So the job has a shape. Turning a scientific question into objectives, constraints, and a bounded search space requires judgment no algorithm currently supplies. The physical instruments need to be kept honest, because unreliable hardware produces data that looks like knowledge and is not. And after each campaign, someone still has to check that the fast automated measurements correspond to the physical structures anyone cares about. The automation occupies the middle of that sequence, running inside boundaries the researchers spent months defining.