Efficient Exploration of Multimetallic Alloys for Optimal Bifunctional Catalysts in Water Splitting through Pareto Active Learning and Experiments

The search for efficient bifunctional multimetallic alloy catalysts is a significant issue for the production of renewable energy, particularly in water splitting. In a recent study published from Prof. Hee-Tae Jung’s group, an optimal bifunctional catalyst for water splitting was obtained by combining Pareto active learning and experiments. The study focused on the use of multimetallic alloys, which have large dimensions and complex compositions, making the search for high-performance catalysts challenging using conventional trial-and-error experiments.

The researchers utilized a Pareto active learning framework and carbothermal shock (CTS) method for nanoparticle synthesis to determine the optimal metallic compositions for bifunctional catalysts with up to four component elements for hydrogen evolution reaction (HER) and oxygen evolution reaction (OER). The Pareto active learning framework was built on two Gaussian process regressors (GPRs) that predicted the overpotential and evaluated the uncertainties of their own predictions, suggesting the composition for the next round of experiments to be conducted. The models reduced the uncertainty of the model for HER and OER overpotential values, and the search trajectory revealed consistent improvements during the iterations, resulting in catalysts with high catalytic properties, exhibiting a cell voltage of less than 1.6 V at a current density of 10 mA cm−2.

This study demonstrated that by appropriately constructing models with a small amount of experimental data, high-performance catalysts can be discovered using the Pareto active learning framework. The approach can be extended to different catalytic reaction fields and other multifunctional applications, such as hydrogen storage, supercapacitors, and photocatalysts. One possible direction for further research includes introducing problem-specific learning models to enhance the performance of the prediction models in the active learning loop, either by feature engineering, kernel optimization, or increasing the complexity of the models.

This work was supported by the Saudi Aramco-KAIST CO₂ Management Center.