**HIGH THROUGHPUT METROLOGY FOR BOTH ACCURACY & PRECISION**

High magnification is not necessarily the best strategy for nanoparticle metrology that is both accurate and precise. In this study the nanopillars on a cicada wing were imaged using secondary electrons with a through-lens detector usings magnifications from 40KX to 400KX in 10KX increments to allow for a continuously increasing number of target objects. The number of nanopillars imaged covered three orders of magnitude from 16 to 1792 subjects, all imaged with the same focus, working distance, and number of image pixels. NanoMet was used to perform metrology on the diameter of the pillars with simple Otsu thresholding and no image preprocessing beyond pixel size calibration using the image length bar. Imaging started at 400KX and then the magnification was reduced in 10KX increments without any adjustment to the image shift or stage position, allowing the target objects in the high magnification images to be a subset of those imaged at low magnification.

Nanopillar diameters and standard deviations were measured using NanoMet and plotted as a function of the number of nanopillars measured. Both metrics fluctuate considerably for images representing less than 100 nanopillars. The fluctuations of these metrics are not statistical in nature. As the magnification decreases and more subjects are measured by NanoMet, a larger sampling of a non-uniform distribution of nanopillars occurs.

While one needs a sufficient number of measured objects to reduce statistical variation and thus achieve low measurement uncertainty– one also needs a sufficient number of measured objects in the sample to accurately represent the object population without introducing measurement bias. The low magnification image of the cicada wing shows discrete domains of nanopillars with significant spatial variation in the size of the nanopillars. There is a mosaicity to the image with nanopillars ordered in domains with significant size variation occurring between domains.

At high magnifications there was a sufficient statistical sampling to allow for nanometer measurement uncertainty, less than 1% of the average diameter. However because the variation in the whole population of nanopillars was not being sufficiently sampled, the mean diameter at high magnifications had low statistical uncertainty, but poor accuracy. It took at least 200 measured

Magnifications allowing NanoMet to measure at least 200 objects (110KX and below) provided measurements of the nanopillar diameter that had low statistical uncertainty but which were also accurate. It is a different matter if one considers the diameter range because the smallest and largest diameter pillars are fewer in number, and subject to the largest statistical variation. Similar variations were seen in the mean diameter, but here the diameter range is accurate with at least 800 nanopillars included in the NanoMet measurement run.

One systematic source of these fluctuations in diameter as measured by NanoMet are the measurement of nanopillars at the edges of image fields. NanoMet allows the operator to define a confidence interval that defines the fraction of a full circle that must be fit around an object to be considered a particle. In this study that confidence interval was set to 75% requiring three quadrants of a circle to be fit to each nanopillar. It is useful to include particles at image field edges, particularly at high magnifications, as these particles represent a large fraction of the particles in those high magnification images. However small changes in magnification change the amount of an edge particle present in the image field, and thus the estimate of the diameter of those nanopillars.

While the specifics will vary depending upon the sample and its particular distribution of particle sizes and how they are distributed in their corresponding SEM or TEM images, in the case of this cicada wing, at least 800 measured nanopillars are required to characterize the nanopillar diameter mean, variation, and range. In terms of automatic metrology run-time, NanoMet was able to measure and generate those metrics in less than one minute using a single 60KX image field. Alternately, NanoMet can batch run images that contain fewer objects. As an example NanoMet could perform the same characterization on twenty-seven 300KX images in about two minutes of run time. Neither of these approaches are possible with a significant amount of human interaction with these images, making NanoMet the tool for precision, reliability, and accuracy.