Optimizing Molecular Descriptors for Reliable Adsorption Energy Prediction on Transition Metal Nanoclusters

Efficient catalytic processes are crucial for converting pollutant molecules into valuable products. Transition-metal nanoclusters show promise as a result of their tunable properties, but identifying active catalysts requires costly studies of the adsorption energetics. Machine learning offers a faster alternative, predicting adsorption energies when trained on suitable descriptors and relatively large density functional theory (DFT) data sets. This study evaluates the predictive power and transferability of two structural descriptors, the Coulomb matrix and the many-body tensor representation, on a diverse nanocluster-adsorbate data set using the random forest regression algorithm. Both descriptors achieved a mean absolute error of 0.05 eV in test data, but performance dropped significantly on an external generated set with unprecedented examples. Adding a simple electronic feature, the number of unpaired electrons of the adsorbate improved generalizability, even though with higher mean absolute errors compared to the original data set, highlighting the dependence on the training data.