Towards Machine-Learning Enabled Automated Analysis of Heterogeneous Nanomaterials

This is the authors' abstract. We don't add key points for this paper.

Automation is playing an increasingly important role in the discovery, design, and synthesis of materials. In the field of nanomaterials, high-throughput robotic solution phase and thin film synthesis are increasingly adopted to optimize nanomaterial properties and explore design spaces [1, 2]. Similarly, many experimental characterization methods are increasingly automated, often with the goal of combing these measurements with robotic synthesis tools to perform closed-loop synthesis. Given the important role that local, atomic-scale features play on nanomaterial properties, transmission electron microscopy (TEM) is heavily utilized to relate structure to properties in these systems. TEM provides a variety of information at different length scales related to a material’s structure and composition. For example, a single image from an aberration corrected electron microscope can measure atomic positions, defects, and strain with sub-Angstrom resolution. Recent advances in electron detection have enabled routine use of computational imaging methods based on scanning nanodiffraction, such as 4D-STEM and ptychography. Taken together, these advances in detector speed, automated imaging, and robotic sample preparation strongly motivate tools for automated analysis of electron microscopy data. Machine learning approaches for data reduction and interpretation have also been the subject of intense recent study, particularly in the field of image segmentation and classification. Convolutional neural nets (CNNs) have emerged as a popular choice to perform these tasks, and recent work has demonstrated the utility of CNNs as part of high throughput image analysis pipelines [3, 4]. In these examples, CNNs are used to identify nanoparticles and automatically determine size and shape statistics. This approach is extremely useful to determine how nanoparticles evolve during synthesis or use, and to relate the size and shape statistics of a nanoparticle population to bulk properties. Previous work has illustrated that CNNs require special consideration to avoid bias when implemented for TEM data analysis, and may not generalize well when experimental conditions change [5, 6]. They also require training sets that adequately capture the feature space of the data to be analyzed, which presents a challenge for automated imaging experiments where unexpected features may be present. In particular, the degree of sample heterogeneity plays an important role: to obtain adequate statistical sampling of images of nanomaterials, the number of observations needed varies depending on the variety of features present. This is captured by the concept of ‘representativeness’, a property that describes whether a subset of data that captures the properties of the entire dataset. Even for seemingly self-similar populations of nanomaterials, large sample sizes are needed to observe subtle trends in size and shape statistics [7]. This work illustrates implementation of machine learning approaches for automated analysis of large nanomaterial datasets, with an emphasis on statistical sampling, proper curation of training data, and strategies to avoid bias in analysis. Ultimately, this work highlights important considerations for automated data analysis when sample or experimental conditions may be changing or unknown [8].