Construction of Machine Learning Potentials Toward the Exploration of High Entropy Alloy Cluster Catalysts

In recent years, high entropy alloys (HEAs), which contain at least five elements, have received considerable attention due to their unique properties that make them promising candidates for a variety of applications [1]. However, searching for the best HEA to serve a purpose within the huge compositional space by first-principles calculations based on density functional theory (DFT) is computationally expensive. Therefore, we constructed machine learning potentials (MLPs) to efficiently search for HEA catalysts.<br/>In this study, we aim to identify HEA catalysts for the carbon dioxide reduction reaction. As the activity and selectivity of this reaction are highly correlated with the adsorption energies of carbon monoxide (CO) molecules and hydrogen (H) atoms [2], MLPs were trained to predict these values. The training data were obtained from DFT calculations of models with different atomic configurations in bulk crystals (8 atoms per supercell) containing up to 5 of the 9 elements (Co, Ni, Cu, Rh, Pd, Ag, Ir, Pt, and Au), and in octahedral clusters (19 atoms) containing up to 4 of the 9 aforementioned elements. Here, the results for clusters were included to investigate the catalytic performance not only on flat planar surfaces but also on uneven surfaces.<br/>Our MLPs were constructed using a graph neural network-based architecture with the Allegro package [3] and trained with the DFT data. Our results show that the MLP accurately predicts the total energies of the quaternary and quinary alloy models from the training data of the binary and ternary alloy models with a mean absolute error (MAE) of approximately 12 meV/atom. We also confirmed the diversity of our dataset across the entire chemical component space by examining the distribution of high-dimensional features representing the local atomic environment via dimensionality reduction methods (i.e. t-SNE and PCA). In addition, the adsorption energies of H atoms on the octahedral clusters (19 atoms) of binary and ternary alloys were predicted with a MAE of approximately 0.1 eV. In conclusion, it is promising to adapt our machine learning potential in designing high-performance HEA catalysts for carbon dioxide decomposition.<br/>This study was supported by JST e-ASIA Joint Research Program.<br/><br/>[1] M.-H. Tsai and J.-W. Yeh, <i>Mater. Res. Lett.</i> <b>2</b> (2014) 107.<br/>[2] J. K. Pedersen, et al., <i>ACS Catal.</i> <b>10</b> (2020) 2169.<br/>[3] A. Musaelian, et al., <i>Nat. Commun.</i> <b>14</b> (2023) 579.