Unsupervised Learning to Understand the Structural Transformation of Ultrathin AuAg Nanowires into Double Helical Structures Using 4D-STEM

There have been several literature reports of ultrathin (&lt; 2 nm diameter) colloidally grown, Au-based nanowire systems twisting into double helices upon induction with either Au-, Pt,-, or Pd- based metal salts. The phenomenon that allows these single nanowires to twist into double helices is not well understood and conflicting arguments in the literature have been presented. Developing an understanding of the underlying forces that govern this twisting behavior could allow for improved design of ultrathin 1D chiral metamaterials and thus improvements in catalytic performance. We focus our analysis on the evolution of 1D seed nanowires (&lt;2 nm AuAg) into either straight or twisted core-shell Palladium-on-Gold/Silver (Pd@AuAg) 1D nanowires with a diameter of approximately 5-10 nanometers. In particular, we explore the structural differences between nanowires that undergo twisting and those that do not using diffraction imaging.<br/>Four-dimensional Scanning Transmission Electron Microscopy (4D-STEM) has become widely popular due to the ability to resolve structural information of complex materials at the sub-nanometer scale. Orientation, phase, and strain mapping methods all provide crucial information towards understanding dynamic processes in the synthesis of these twisted 1D nanostructures. Additionally, the advent of new electron detector technologies has allowed for several thousand diffraction patterns to be collected in a matter of minutes, allowing for great strides to be made in sub nanometer structural characterization of materials.<br/>As datasets increase in size, intelligent workflows to compress these datasets into a smaller set of user-interpretable features that capture the spirit of the underlying images are crucial steps towards storing and processing the rich sets of structural information within these patterns. Current available workflows focus primarily on the reduction of these complex patterns to the locations and intensities of the Bragg disks, which are not representative of all material structures. There has been a deficiency of workflows compatible with understanding diffraction information of disordered materials since the structural information is not entirely captured in workflows that are only sensitive to Bragg disks.<br/>The feature extraction workflows we explore to gain more detail on the twisted nanowire system are related to angular and radial intensity distributions within the diffraction patterns. We use the angular average and root mean square deviations (RMSD) to supplement information captured in the widely used Bragg disk detection workflows. We demonstrate that the use of unsupervised learning can capture disorder at the grain boundaries within 1D materials, which is not detected using the Bragg disk data representation method. These additional representations allow more decisive classification of disordered regions within the sample that are missed with other methods. By applying this newly developed feature extraction workflow, we can better discern structural differences within and between populations of 1D nanowires. We expect these workflows to be broadly applicable to ultrasmall (&lt; 10 nm) 0D and 1D materials, which typically contain significant disorder as opposed to their bulk counterparts.