Prediction of Pd, Pt, and Rh adsorption energies on graphene quantum dots using graph neural networks

LARA GONCEBAT; BELLETTI, GUSTAVO DANIEL; PAOLA QUAINO; RODRIGO ECHEVESTE ; MATIAS GERARD

Tennesee

Conferencia; 26th IUPAP Conference of Computational Physics; 2025

Oak Ridge National Laboratory

Transition metals such as Pd, Pt, and Rh exhibit excellent electrocatalytic activity but sufferfrom the disadvantages of being expensive and scarce. An effective strategy to optimize theiruse is to employ them as adatoms on suitable supports. In this context, graphene quantumdots (GQDs) are notable due to their abundance, large surface area, the presence of active edgesites, and interesting geometric structures, making them promising supports for electrocatalystdesign. The interaction between these carbon-based systems and transition metals has shownpromising results in improving stability and catalytic activity [1]. At the same time, artificialintelligence (AI) and machine learning (ML) are transforming materials design by enabling theefficient exploration of vast chemical and structural spaces, with computation times significantlyshorter than conventional methods [2]. Within this context, graph neural networks (GNNs)have emerged as particularly suitable tools for modeling non-periodic atomic systems, as theyexplicitly capture connectivity, local interactions, and spatial symmetries without relying onperiodic cells. Their application in materials science allows property prediction at substantiallylower computational cost compared to conventional approaches [3]. In this work, densityfunctional theory (DFT) calculations are combined with GNNs to predict the adsorption energiesof Pd, Pt, and Rh on GQDs of different geometries. A database of adsorption configurationswas generated, incorporating electronic, geometric, and charge properties. Each system wasrepresented as a graph,with nodes corresponding to atoms and edges reflecting connectivitywithin a cutoff radius. As input attributes, we incorporate the previously calculated propertiestogether with relative coordinates, atom type, and local chemical environment, which allows amore accurate representation of the relationship between structure and properties. Based onthese representations, GNN models were trained and evaluated using specialized graph deeplearning libraries. These networks update node states through a message-passing scheme inwhich each atom aggregates and transmits information from its neighbors to capture local environmental effects. For training, we employed MSE loss functions, global pooling strategies toobtain system-level descriptors, and graph convolution–based architectures, complemented bycross-validation to assess generalization. Preliminary results demonstrate good fitting and strongpredictive performance, highlighting the potential of GNNs to accelerate property prediction andguide the rational design of new electrocatalysts. [1] Peng Y., Lu B., Chen S., ADV MATER,2018, 30, 48, 1801995. [2] Mai H., Le T., Chen D., Winkler D., Caruso R., CHEM REV, 2022,122, 16, 13478-13515. [3] Reiser P., Neubert M., Eberhard A., Torresi L., Zhou C., Shao C.,Metni H., Van Hoesel C., Schopmans H., Sommer T., Friederich P., COMMUN MATER, 2022,3, 93.