Electron Microscopy: Deep Learning-Based Autofocus and Astigmatism Correction

Modern electron microscopy demands high precision in focusing and astigmatism correction due to increased resolution and sample complexity. Existing automatic methods often rely on mathematical models of specific microscopes, requiring deep domain knowledge and extensive computation. These limitations make them slow and inflexible, especially when imaging large or diverse samples. Furthermore, their lack of general applicability hinders broader adoption. As imaging needs scale in terms of both volume and precision, a universally adaptable and efficient optimization tool becomes essential. The present invention addresses these shortcomings with a data-driven, learning-based solution that performs fast, accurate adjustments across multiple setups.

The invention introduces a deep learning-based method for the automatic adjustment of imaging parameters in electron microscopy, specifically focusing on working distance and astigmatism correction. The core innovation lies in replacing traditional model-based optimization routines with a data-driven approach that is fast, generalizable, and easy to integrate.

In the optimization process, two images are acquired from the microscope with known perturbations applied to the current suboptimal settings. These image pairs are divided into multiple subregions (patches), which are then processed by a convolutional neural network trained to estimate correction values for each image patch. The network outputs predicted adjustments for the working distance and both x- and y-components of astigmatism. These individual predictions are then aggregated to calculate a single, robust correction vector for the overall image.

After applying the calculated correction, a new image is taken, and the process is repeated until optimal image quality is achieved - typically within three iterations. The system requires only standard computational hardware (CPU with memory), making it suitable for integration into existing microscopy platforms.

The method is adaptable to different microscopes and imaging tasks, and its modular structure allows for the inclusion of additional adjustable parameters. It enables real-time, automated optimization with minimal user intervention and high reproducibility across diverse imaging conditions.

• Ready-to-use solution: Easily integrates into existing microscope systems without hardware changes.
• Single environment-specific training: One-time training adapts the AI to specific microscope setups.
• Wide compatibility: Works across different microscope types, models, and parameter sets.
• Fast and efficient: Provides reliable correction results within just a few iterations and minimal processing time.
• Scalable and flexible: Accuracy and scope can be extended by adjusting input image volume or applying the method to additional imaging parameters.
• Minimal user input: Once trained, the system operates autonomously, reducing operator workload.

• Scanning Electron Microscopy (SEM): Improves focus and distortion correction in standard SEM applications.
• High-dose imaging: Maintains focus precision under strong electron beam conditions.
• Automated 2D/3D imaging: Integrates into high-throughput and robotic acquisition pipelines.
• Large-sample imaging: Optimizes focus and astigmatism in wide-field or high-content microscopy.
• Precision research and diagnostics: Useful in disciplines requiring sharp, consistent image quality, such as materials science or biomedical research.

PCT application (WO2023099584A1) (19.01.2022)