Choosing a plate, not just an experiment

Imagine a plate of reaction vessels ready to run together. A useful planning question is whether every vessel should chase the same promising region, or whether some should test alternatives. For our launch methods series, Minerva provides a concrete case for examining that choice: a 2025 study combines automated high-throughput experiments with machine-learning selection of parallel reaction batches.¹

Two objectives on one plate

The authors apply the workflow to a nickel-catalysed Suzuki coupling in 96-well batches, tracking both area-percent yield and selectivity.¹ Those two measures make the decision more interesting than simply finding the largest product signal. Our reading is that an optimisation report should show the trade-off offered to the chemist, rather than silently collapsing every useful outcome into a single winner.

The experimental campaign begins with systematic Sobol sampling, followed by Bayesian optimisation; the allowed chemical search space was defined by chemists.¹ In other words, the algorithm's freedom has an explicit boundary. When assessing another system, we would ask to see that boundary before comparing its performance: which choices were permitted, which were excluded, and why?

The measurement label matters

The reported reaction search reaches up to 76% area-percent yield and 92% selectivity.¹ The methods explain that the yield proxy comes from liquid-chromatography signals without correction for differences in detector response between product and limiting starting material.¹ We therefore retain the area-percent label. It would be misleading to present the number as a weighed, isolated yield.

For a reader following a reaction towards practical use, this distinction suggests a useful handover document. Put the screening measurement beside the later measurement intended to settle the process decision. State what each one answers. That document should make disagreements visible rather than make an early screening number carry a stronger meaning than it was designed to support.

A human changes the search

Late in the campaign, the researchers deliberately favour exploitation of promising conditions over further exploration, changing the acquisition strategy for the final iteration.¹ This is an informative intervention. The observed result belongs to a workflow with an expert decision inside it, rather than to an algorithm left entirely alone.

Our preferred comparison would preserve a record of that decision and the alternatives available at the time. It would also distinguish the benefit of the first structured sampling plate from the benefit of later learning. Calling all improvements an AI effect would leave the reader unable to tell which component earned its place.

What we would carry forward

The practical lesson we draw is about reporting a campaign as a sequence of accountable choices. Keep the objective definitions, plate constraints and intervention history next to the outcome. A strong screening result then becomes a reason to investigate a condition further, with its evidence boundary intact.

This article revisits a published experimental study; it is not a newly performed optimisation or an independent reproduction. We would judge a subsequent deployment by its own complete campaign record, including failed wells and the work needed after the screening stage.