A robot lab tested 1,200 catalyst recipes with an interpretable AI controller

An automated laboratory tested about 1,200 catalyst recipes over roughly 3,000 hours, using an AI controller designed to make the search interpretable. The Nature Communications study, published on 23 October, reports promising hydrogen-evolution activity from earth-abundant materials and a model that lets researchers examine how nearby synthesis choices affect performance.[1]

The controller is a chemistry-infused bandit: it chooses the next experimental setting, receives the measured result and updates its search. Automated synthesis and electrochemical testing supply the feedback. The aim is to learn about the chemical response while looking for a better catalyst.[1]

Twelve knobs on one chemistry problem

The campaign explored a 12-dimensional space of synthesis conditions for the electrocatalytic hydrogen evolution reaction. Each setting can affect the structure and activity of a material. Searching that space with traditional one-factor-at-a-time experiments is difficult, while a purely black-box optimiser may find a strong sample without supplying a practical account of why it performs well.[1]

The model encodes a chemist’s hypothesis: near a promising region, activity can be approximated by a local quadratic relationship with synthesis conditions. This lets it represent curvature and interactions between variables while keeping parameters that scientists can inspect. The approximation is local; it is not a universal law of catalyst behaviour.[1]

A bandit balances trying uncertain options against testing settings already expected to perform well. Here, the experimental reward reflects hydrogen-evolution performance. An interpretable response model gives the team a way to inspect why a region looks promising and propose chemical explanations for the pattern.

What the experiments established

The reported overpotentials were 39 millivolts at 2 milliamps per square centimetre and 198 millivolts at 50 milliamps per square centimetre. Overpotential is the extra voltage required beyond the reaction’s equilibrium potential; at a specified current density, lower values indicate better catalytic activity. Reporting both current and voltage makes the performance claim more informative than saying only that a catalyst rivals platinum.[1]

The comparison concerns the study’s alkaline laboratory tests. Operating lifetime, catalyst loading and full-cell behaviour also affect whether a material is suitable for an electrolyser. The measured activity is a lead for further development, rather than evidence that the resulting system can replace a commercial hydrogen-production process.[1]

Why an interpretable controller is different

The model’s value extends beyond selecting the strongest sample. Its response patterns give chemists hypotheses about influential variables and interactions near the explored settings. Those hypotheses can guide further experiments, provided the team checks that the local approximation still applies when it moves into a new region.[1]

The campaign connects three pieces that are often separated: choosing experiments, measuring activity and interpreting the response. Roughly four months of automated testing produced both candidate catalysts and a structured account of where to investigate next. The quality of that account is part of the scientific output.[1]

Report an error or suggest a correction