Machine Learning for Structural Characterization and Generation: Applications to Small-Angle Scattering and Electron Microscopy
The era of big-data-driven science has brought to light the need for new methodologies to process and extract otherwise-unattainable insights from the vast amounts of data generated by materials and nanostructural characterization methods. Small-angle scattering (SAS) and electron microscopy (EM) experiments yield rich data sets that contain structural and morphological information of nanostructures. Exploiting data-driven methods to extract these insights is a natural fit. Additionally, the volume of data produced in the previous computational and simulation age of science has led to the establishment of extensive repositories of structures. These resources present opportunities for functional-property prediction to reveal novel uses for existing structures and for deep generative models to design new structures for a wide range of applications.
This thesis is concerned with the development of machine learning (ML) algorithms for the characterization and generation of materials and nanostructures. Chapter 1 discusses characterization and generation in the data-driven age of science and reviews the application of ML to aid these processes, with a focus on SAS and EM for characterization. Chapter 2 provides a technical outline of ML in general, including the specific methods that are employed in subsequent chapters to develop models for processing SAS and EM data for characterization, as well as atomic structures for property prediction and generation. In Chapter 3, a convolutional neural network-based segmentation algorithm is developed to detect and locate nanoparticles in EM images. This constitutes the particle segmentation module of ImageDataExtractor, an open-source software tool developed therein for extracting information from EM images in an automated fashion. Techniques from Chapter 3 are employed in Chapter 4 to calculate SAS intensity functions from morphological and position information of nanostructures obtained from 2-dimensional EM images. In Chapter 5, the focus shifts to SAS, where a multi-task neural network is developed to concurrently identify the size and shape-parameters of the scatterers that produced a given SAS intensity. Chapter 6 constitutes the property prediction and structural generation portion of this thesis, in which a deep generative model of inorganic crystal structures is developed alongside a graph neural network to predict the properties of the generated structures. Finally, concluding remarks and avenues for future research are discussed in Chapter 7.