Geometric Guidance Integrated with Directed Electrostatics Strategy within a Graph Neural Network Approach for Nanocluster Structure Prediction

We introduce the Geometric-DESIGNN method, which integrates Geometric Guidance with Directed Electrostatics Strategy within a Graph Neural Network framework to predict the stable configuration of nanoclusters on their potential energy surfaces. This approach merges the geometric and electronic strategies using graph neural network-based models to predict structures of large atomic clusters with specific size and point-group symmetries. This approach aids in constructing atomic metal cluster structures by predicting building frames through a geometric approach and locating the minima in the molecular electrostatic potential (MESP) landscape. By following alternate geometric and DESIGNN building strategies for each shell of parent clusters, we efficiently achieve close-packed daughter structures along their evolutionary paths. The geometric-DESIGNN approach is validated on the prototype Mg_<i>n</i> clusters, by building structures for sizes up to <i>n</i> < 561. Furthermore, constraining the point-group symmetry of the parent clusters, we identify new symmetric isomers of medium to large Mg_<i>n</i> clusters with <i>n</i> < 150. This methodology is also employed to construct stable Mg_<i>n</i> nanoclusters with <i>n</i> = 332, 338, and 561. Benchmarking results show that the geometric-DESIGNN approach is an efficient tool for accelerated prediction of the nanocluster structure.