High-throughput microscopy techniques, including scanning electron microscopy (SEM), have become increasingly important for the time-efficient analysis of large volume samples in materials science Lang et al., 2024. SEM is an important method in microstructural characterization, because it offers high spatial resolution at the nanometer scale, thereby enabling detailed characterization of microstructural heterogeneities, identifying features such as grains, defects, and textures across various length scales, depending on the employed technique Reimer, 1998,Goldstein et al., 2018. Additionally, SEM is well-suited for high-throughput approaches due to its flexibility in selecting various scanning parameters and fast read-out of detector signals Goldstein et al., 2018.

In SEM, a focused electron beam scans across a sample’s surface, resulting in the emission of secondary electrons (SE) and backscattered electrons (BSE). Low energy SEs are generated by inelastic scattering events and are generated within a few nanometers of the sample surface, and are therefore highly sensitive to surface topography Reimer, 1998. On a well-polished sample surface, SE imaging, therefore, provides important details on structural integrity as well as structural defects at the micrometer level. However, so-called edge effects introduce challenges in SE image analysis, where features at the edges of structures appear brighter or distorted due to changes in the electron interaction volume and angular emission at boundaries Reimer, 1998.

Optimizing image acquisition time relative to the required image resolution for resolving structural features, several acquisition parameters must be balanced, such as high tension and beam current, but also - for a given high tension and beam current - dwell time and pixel size. Dwell time, the time that the electron beam remains on each pixel, plays a critical role in determining the signal-to-noise ratio (SNR) of the acquired images Joy, 2008. While shorter dwell times speed up the data acquisition process, they tend to introduce more noise, which can decrease feature detection accuracy Rahman et al., 2024. On the other hand, longer dwell times improve SNR, but they significantly increase the overall acquisition time, which limits SEM’s utility in high-throughput workflows. Pixel size is another essential parameter; smaller pixel sizes are beneficial for a higher spatial resolution, while larger pixel sizes further reduce the acquisition time, but will limit image resolution, i.e., the ability to resolve a specific feature such as a small defect with a size at or below the pixel size. Fortunately, SEM images can nowadays routinely be acquired sequentially from neighboring areas and stitched together to capture a larger field of view, overcoming the traditionally limited field-of-view in a single SEM image Lang et al., 2024. Fine-tuning the aquisition parameters however is challenging because the optimum often depends on the specific characteristics of the sample being analyzed.

Automated feature detection workflows are becoming increasingly prominent in quality control procedures for a wide variety of materials Casukhela et al., 2022 with the goal of reducing inconsistencies based on human input and improving precision Diers & Pigorsch, 2023. While significant advancements in using high-throughput techniques for feature detection have been made, manual steps still present in these methods remain time consuming and error-prone Dehaerne et al., 2024. Various studies have explored automated solutions to address these challenges, integrating machine learning (ML), Bayesian optimization, and image processing algorithms into SEM workflows Casukhela et al., 2022. Many of these automated systems focus on either defect detection or image acquisition optimization without addressing either of them comprehensively. For instance, ML models have been applied to enhance defect classification accuracy, but these often require high-quality images, which increase acquisition times Lobato et al., 2024. On the other hand, efforts to reduce acquisition times typically lead to a loss in detail that affects defect detection accuracy and statistical relevance Rahman et al., 2024.

We selected additively manufactured (AM) steel samples as a model system to investigate the accuracy in materials’ defect detection and to optimize SEM acquisition parameters to obtain reliable statistics on such defects. Metals are well-suited for such a study since they remain stable under electron beam exposure, allowing repeated analysis of the same area under varying acquisition conditions. Furthermore, AM components are often several centimeters in size, while critical microstructural features, such as grains, phases, and defects, are in the micrometer range. Defects like pores and cracks can significantly impact the mechanical performance and durability of AM components Ellendt et al., 2021, and their morphological characteristics and distribution within samples provide valuable insights into their origins Casukhela et al., 2022, Ellendt et al., 2021. This underscores the importance of large-area scans and optimized SEM acquisition times to capture essential microstructural details accurately across varying scales.

Here, we present a Python-based framework for analyzing features in large-area SEM micrographs, utilizing local contrast conditions to improve detection accuracy. This framework allows for a detailed analysis of how dwell time and pixel size affect feature identification’s accuracy and detection limits by comparing results to a “benchmark” established by high-quality SEM micrographs acquired with long dwell time and small pixel size. Finally, we explore a noise reduction (i.e. denoising) method to recover detection accuracy under low dwell time, low SNR conditions, providing a path to optimize SEM imaging for both efficiency and precision in defect analysis.