GIFT Yujun Xie's Team Publishes in npj Computational Materials: A unified prepossessing AI Framework for Materials Analysis

A research team led by SJTU Global Institute of Future Technology Associate Professor Yujun Xie has published a paper titled “A unified preprocessing framework for high-throughput diffraction pattern analysis” in the journal npj Computational Materials. To address the critical bottleneck in high-throughput four-dimensional scanning transmission electron microscopy (4D-STEM), the team proposed a unified deep learning framework "4D-PreNet" to streamline the preprocessing of complex microscopy data. The technology simultaneously performs denoising, beam center calibration, and elliptical distortion correction on diffraction patterns. It paves the way for automated, real-time analysis of material structures and holds significant value for the large-scale application of microscopy analysis techniques in engineering and industrial scenarios.

4D-STEM, which captures a two-dimensional diffraction pattern at every scan position, has emerged as a powerful technique for mapping local crystal orientation, strain fields, and defects in materials. However, its practical implementation is limited by challenges in data preprocessing. Issues like pervasive noise, beam center drift, and elliptical distortions during high-throughput acquisition inevitably degrade diffraction patterns, leading to systematic errors in quantitative measurements. While existing methods have integrated various conventional calibration algorithms into community tools such as py4DSTEM, they typically require manual parameter adjustment for each dataset and fail to provide a robust, generalizable solution.

Figure 1: Deep learning framework of 4D-PreNet for 4D-STEM preprocessing. a) Overview of the end-to-end three-stage deep learning pipeline for 4D-STEM data preprocessing. b) Detailed workflow illustrating denoising, center detection, and ellipse calibration networks with intermediate outputs.

To address the challenge, the team proposed “4D-PreNet”, an end-to-end deep learning pipeline that integrates multiple previously separate prepossessing steps - including denoising, beam center calibration and ellipse calibration - into a single workflow. This integration reduces reliance on manual expertise and complex parameter tuning while maintaining stable performance across different material systems and experimental conditions. It provides a novel solution for automated, standardized processing of 4D-STEM data.

The framework is trained on a large-scale simulated dataset covering diverse crystalline and amorphous materials. Researchers generated training samples that included complex factors such as noise, drift, and distortion, enabling 4D-PreNet to learn stable representations of diffraction patterns across different materials and imaging conditions. This approach equips it with strong cross-material and cross-scenario generalization capabilities. 4D-PreNet demonstrates robust performance across different material systems and experimental conditions without the need for manual tuning. It represents a significant step toward transforming 4D-STEM from a high-throughput acquisition tool into a platform capable of automated, near-real-time analysis.

Figure 2: End-to-end preprocessing of experimental 4D-STEM data using 4D-PreNet. a) Representative diffraction patterns from amorphous, crystal, and mixed regions. b) Center calibration results before and after calibration, with improved beam alignment across amorphous, crystal, and mixed regions. The right panels display the corresponding X and Y shift maps, showing reduced spatial drift after calibration. c) Ellipse calibration applied to averaged diffraction patterns. Both the average and polar-transformed profiles demonstrate improved circularity of diffraction rings, confirming effective calibration of elliptic distortion.

Mingyu Liu, a GIFT master's student, is the first author. Associate Professor Yujun Xie, Assistant Research Professor Shufen Chu, Professor Xiaoqin Zeng from the School of Materials Science and Engineering, and Zian Mao from SJTU Global College are co-corresponding authors. The research exemplifies a successful industry-academia partnership, advancing the framework's industrial applicability and paving the way for the deployment of high-throughput microscopy in corporate R&D. The work was supported by the National Natural Science Foundation of China.

Paper Link:

https://doi.org/10.1038/s41524-026-01993-3

Author Profile

Mingyu Liu

Master's student(enrolled in 2024) in Electronic Information, Global Institute of Future Technology, SJTU. Research interests: 4D-STEM data processing and automated analysis. He has been honored with several awards, including the AMEC Future Scholarship, and the First-class Academic Scholarship, and holds a software copyright.