image: Spatial-adaptive active-learning workflow discovers long-lasting acidic OER catalyst with 625-hour durability
Credit: ©Science China Press
Researchers in Prof. Tong-Yi Zhang’s lab have introduced a spatial-adaptive active-learning workflow that significantly accelerates the search for highly durable OER catalysts, addressing one of the toughest challenges in acidic water electrolysis for green hydrogen production. Discovering catalysts that simultaneously exhibit low overpotential and long-term stability is hindered by the time-consuming nature of stability testing, even though overpotential can be measured rapidly. This imbalance makes conventional exploration prohibitively slow and costly.
The new workflow solves this bottleneck by optimizing the two objectives sequentially within a unified active-learning framework. In the first stage, Bayesian optimization is used to identify low-overpotential regions and characterize how overpotential is distributed across the compositional feature space. In the second stage, a conditional variational autoencoder (CVAE) is applied to rapidly generate a refined low-overpotential subspace of 785 candidates. The search is then confined to this subspace, where the algorithm focuses exclusively on durability optimization. With only three iterative cycles involving three additional synthesized samples and eleven long-term stability tests, the workflow identifies a Cu-doped RuO2 catalyst with exceptional performance.
The discovered Cu-RuO2 catalyst delivers an overpotential of 177 mV at 10 mA cm−2 in 0.5 mol L−1 H2SO4 and operates continuously for 625 h, representing one of the longest stability records reported for acidic OER catalysts under comparable conditions. This study demonstrates how combining machine learning with closed-loop experimentation can substantially reduce the need for lengthy validation tests and accelerate the discovery of catalysts for industrial water electrolysis. The researchers note that the method is generalizable to other multi-objective materials systems where validation cost is highly unbalanced across different targets.