Fine control of nanocrystal shape is crucial to tailor the functionality of individual particles as well as to target and optimize their collective phenomena. It has been demonstrated that nanocrystal morphology has a direct impact on applications, including catalysis, drug delivery, optoelectronics, and plasmonics. Despite the significance of particle morphology, existing synthetic methods to gain morphological control and elucidate the synthesis-structure-property relationships are complex. New developments in automated particle morphology classification and quantification could greatly advance these investigations, but implementation into machine learning models still relies on the limits of human accuracy for ground truth comparisons, rendering even unsupervised machine learning models to be inherently rooted in human bias. Herein, we use the concept of synthetic image rendering to solve the ground truth problem of particle morphology classification. Simulating the images of a theoretically infinite number of individual nanocrystal shapes allows particle morphology to be defined and labeled exactly by quantitative mathematical functions instead of qualitative labels. The exact nature of such training labels creates an absolute ground truth for the training of a convoluted deep neural network to represent various nanocrystal shapes, eliminating bias altogether. We then demonstrate the accuracy, sensitivity, and practicality of this concept by using this model as the ground truth for an automated pipeline that processes, defines, and classifies experimental TEM images of individual nanocrystals into statistically different morphology groups via unsupervised clustering. Specifically, we compare the morphological classifications of two different syntheses of nominally cuboidal CsPbBr3 quantum dots (i.e., LARP and hot-injection syntheses) done automatically by this pipeline and manually by two different researchers, to assess the effects of human intervention, researcher technique, and synthesis type on resultant particle morphology classification and quantification. Additionally, we show that the pipeline works for other well-defined shapes (i.e., spherical Ni2P nanocrystals) and more morphologically ill-defined systems (e.g., CoNi2S4 nanocrystals), offering an unbiased method for implementing computational techniques in the morphological investigations of a wide range of colloidal nanocrystals.

Creating ground truth for nanocrystal morphology: A fully automated pipeline for rapid and unbiased TEM analysis

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.