A robot laboratory in Poland used reinforcement learning to decide how to investigate reactions, including when to improve a separation before judging the chemistry. Its October paper reports four previously unreported enantioselective transformations after human follow-up.[1]
Why the mirror image matters
Some molecules come in two forms with matching connections between atoms but different spatial arrangements. Like left and right hands, the forms are mirror images that cannot be superimposed. Chemists call such a pair enantiomers. Knowing that a reaction made the expected compound therefore leaves another question: how much of each form did it produce?[2]
The conventional measure, enantiomeric excess, describes the imbalance between the two forms. It is the absolute difference between their fractions, expressed as a percentage. A perfectly even mixture has zero excess; a sample containing only one form has 100%.[3]
For an illustrative calculation, imagine 100 molecules: 95 of one form and five of its mirror image. The excess is 95 minus 5, or 90%. A 90:10 mixture gives 80%. Those figures describe the composition of the pair; they do not tell us how much material the reaction produced. A highly selective reaction and a high-yielding reaction answer different questions.[3]
Choosing the next useful measurement
Two agents controlled reaction screening and chiral analysis. They learned in simulated chromatographic environments, then guided instruments through 672 experiments and 3,439 decisions. They could alter reaction conditions, isolate products or spend another measurement resolving overlapping peaks. The simulations taught decision-making, rather than predicting specific reaction outcomes.[1]
Researchers chose the starting chemical space and selected 18 reactions for catalyst screening. A chemist then repeated promising transformations manually, isolated their products and checked their structures and selectivity. That work established the four new transformations.[1]
The useful advance is the sequence: obtain evidence, decide whether it is interpretable, then choose the next experimental action.[1]
Sources & further reading
- Reinforcement learning guided robotic platform that investigates and discovers asymmetric reactions Nature Communications · 13 Oct 2026 · Accessed 20 Oct 2026
- Basic Terminology of Stereochemistry: chirality IUPAC · Accessed 20 Oct 2026
- Basic Terminology of Stereochemistry: enantiomeric excess IUPAC · Accessed 20 Oct 2026