Machine learning techniques are attractive options for developing highly-accurate analysis tools for nanomaterials characterization, including high-resolution transmission electron microscopy (HRTEM). However, successfully implementing such machine learning tools can be difficult due to the challenges in procuring sufficiently large, high-quality training datasets from experiments. In this work, we introduce Construction Zone, a Python package for rapid generation of complex nanoscale atomic structures which enables fast, systematic sampling of realistic nanomaterial structures and can be used as a random structure generator for large, diverse synthetic datasets. Using Construction Zone, we develop an end-to-end machine learning workflow for training neural network models to analyze experimental atomic resolution HRTEM images on the task of nanoparticle image segmentation purely with simulated databases. Further, we study the data curation process to understand how various aspects of the curated simulated data—including simulation fidelity, the distribution of atomic structures, and the distribution of imaging conditions—affect model performance across three benchmark experimental HRTEM image datasets. Using our workflow, we are able to achieve state-of-the-art segmentation performance on these experimental benchmarks and, further, we discuss robust strategies for consistently achieving high performance with machine learning in experimental settings using purely synthetic data. Construction Zone and its documentation are available at https://github.com/lerandc/construction_zone.

High-throughput scanning electron microscopy (SEM) coupled with classification using neural networks is an ideal method to determine the morphological handedness of large populations of chiral nanoparticles. Automated labeling removes the time-consuming manual labeling of training data, but introduces label error, and subsequently classification error in the trained neural network. Here, we evaluate methods to minimize classification error when training from automated labels of SEM datasets of chiral Tellurium nanoparticles. Using the mirror relationship between images of opposite handed particles, we artificially create populations of varying label error. We analyze the impact of label error rate and training method on the classification error of neural networks on an ideal dataset and on a practical dataset. Of the three training methods considered, we find that a pretraining approach yields the most accurate results across label error rates on ideal datasets, where size and other morphological variables are held constant, but that a co-teaching approach performs the best in practical application.