Summary:

Accurately identifying transition states is an important step in studying catalytic materials. To enable large-scale exploration of transition states in complex materials, we present a method for training machine learning models to predict transition state structures. We applied this method to the formate intermediate, which plays a key role in converting CO₂ to methanol, on copper-supported indium oxide nanoclusters. Our results show that the models predict transition states with high accuracy. We also discovered structure–activity patterns, showing two pathways by which hydrogen can move to the CO₂ adsorbate. Finally, we found that linear scaling relations, traditionally used to screen catalysts, do not hold for these complex materials, highlighting the need for more sophisticated methods like the one outlined in this work.

View paper