Abstract
In recent years, artificial intelligence (A) has entered the field of electrocatalysis as a tool to accelerate electrocatalyst discovery and optimization. Here, we present the full development process of a reinforcement learning-based AI-controller deployed in a closed-loop experimental setup to optimize the glycerol oxidation reaction. The controller autonomously discovered a non-periodic pulsing strategy that adaptively alternates between active and recovery phases. Analysis of the policy and generated data suggests the presence of three distinct deactivation mechanisms: a fast reversible process mitigated by oxidative pulses; a slower reversible process alleviated through near open circuit potential relaxation; and a third, irreversible process leading to permanent deactivation. To support the AI-controller’s development, a data-driven digital twin based on a recurrent neural network, was trained on the experimental data. This digital twin enabled rapid hypothesis testing by simulating the response to user-defined pulsing protocols. Together, this work demonstrates that AI can not only be used to automate experimental design and execution but also to provide mechanistic insights, making AI-driven electrocatalytic research more accessible.

This publication has 22 references indexed in Scilit: