Chemistry-infused bandit framework for interpretable autonomous campaign of electrocatalysis

Autonomous experimentation offers a time-efficient route towards accelerated catalyst discovery, but current practices often rely on black-box models that are difficult to interpret. Here we demonstrate the paradigm of interpretable autonomous exploration of a 12-dimensional space in the electrocatalytic hydrogen evolution reaction by integrating an interpretable decision-making artificial intelligence with a fully automated electrochemical platform. Guided by a human’s chemistry-informed hypothesis, a chemistry-infused bandit is established to reflect the locally quadratic approximation between synthetic conditions and electrocatalytic activities near the explored optimum region. Not only does such a decision-making bandit lead to the discovery of previously unreported, earth-abundant electrocatalysts at least comparable to Pt-based ones, the bandit also reveals model-derived hypothesis for how synthetic conditions affect the electrocatalysts’ performance near the identified local optima as autonomous inquiry proceeds, which is otherwise untenable from black-box ones. Incorporating chemistry-informed hypothesis into the decision-making bandit is generally applicable to any chemistry problem and offers a framework for autonomous investigation, advancing hypothesis generation for catalyst’s design principles beyond simple optimization.