Locating Hydrides in Ligand-Protected Copper Nanoclusters by Deep Learning.

Abstract

Hydrides play an important role in constructing atomically precise metal nanoclusters and nanoparticles. They occupy both the interstitial sites inside the metal cores and the interfacial sites between the surface of the metal core and the ligand layer. Although the heavy-atom positions can be routinely determined by single-crystal X-ray diffraction, the challenge in growing a large and high-enough-quality single crystal for neutron diffraction and the limited availability of neutron sources have prevented researchers from precisely knowing the hydride locations. A recently developed deep-learning method showed great promise in accelerating the determination of hydride sites in metal nanostructures, but it is unclear if this approach, trained on clusters up to Cu32 in size, can be applied to recently discovered, much larger nanoclusters such as Cu81. Here we show that an improved deep-learning model based on convolutional neural networks is both accurate and robust. We apply it to two recently reported copper nanoclusters, [Cu32(PET)24H8Cl2]2- and [Cu81(PhS)46(tBuNH2)10H32]3+, whose hydride locations have not been determined by neutron but were proposed from density functional theory (DFT) calculations. In the former, our CNN model confirms the DFT structure; in the latter, our CNN model predicts a more stable structure with different hydride sites.