Optimization of SEM Instrumental Parameters for Enhanced Imaging Applying Machine Learning

The scanning electron microscope (SEM) is one of the most versatile instruments available for characterizing the micro- and nanostructural properties of solid samples. A major reason is the high resolution that can be achieved when examining bulk objects which makes SEM indispensable in various fields ranging from materials science to biology. Although scanning electron microscopy (SEM) is a well-established technology, there is still room for improvement in both imaging quality and automation. The goal is to achieve an optimal balance between ease of use and performance while minimizing hardware complexity and instrumentation costs, which remain a key advantage of SEM. In this context, the implementation of advanced algorithmic methods for automating acquisition settings is essential to enhance efficiency and reproducibility in analyses. Given the versatility of SEM and its application to a wide range of materials and samples, we have developed a generalized approach that is not limited to a specific sample type, maximizing adaptability and effectiveness across various experimental settings. Instrumental SEM parameters should be carefully optimized for high-quality SEM imaging. The key parameters that influence image quality can be recognized as: Electron beam parameters including acceleration voltage, beam current, and spot size; Geometrical and imaging parameters, which encompass working distance, aperture size, magnification, focus, and stigmation. Detection and image acquisition settings cover detector type, scan speed (dwell time), and drift correction. Sample preparation and environmental conditions include sample coating and chamber pressure. Optimization of the acquisition parameters is crucial for obtaining high-quality images and accurate data. Operator-based adjusting of these parameters presents several challenges, including the interdependence of settings, the risk of sample damage, time-consuming operations, and operator skill dependencies. Moreover, the sample dependency of the acquisition parameters becomes crucial due to the wide range of possible applications of SEM characterization. These difficulties underscore the importance of automation in optimizing operations of SEM settings, making the process more efficient, consistent, and less prone to error. We developed an algorithm-based approach that can identify the optimum conditions for accelerating voltage, working distance, and aperture size for each type of sample. SEM experiments on target sample are conducted based on a full-factorial design in which parameters are varied simultaneously, to explore also the cross-correlation of the considered parameters. To assess image quality, we needed a suitable metric. Since the reference image—obtained under optimum conditions—served as the goal of our investigation, we chose a no-reference image quality metric. The acquired images were analyzed using this metric. A polynomial regression model—a type of supervised learning—was employed as a baseline to determine the optimal operating conditions within the selected SEM parameter space. By predicting the no‐reference image quality metric for any given setting, the model guides the identification of parameter adjustments that minimize NIQE (thus maximizing image quality). Given the small training set, at each new image acquisition, the predicted NIQE is compared with the computed value, and the model is updated with every new data point, continuously refining its predictions and enabling fully autonomous optimization while significantly reducing operator intervention. This approach not only streamlines the SEM imaging process but also significantly reduces the time required for each imaging session by automating the optimization of key parameters. The proposed algorithm promises to enhance the efficiency of SEM operations, minimize operatorintervention, and ensure more consistent, accurate, and faster sample characterization across a wide range of applications.