the Beamline Software Team of HEPS

Figure 1: Frontier Application of AI for Science for Analyzing Massive Data in Multiple Experimental Methods

1. An efficient ptychography reconstruction strategy through fine-tuning of large pretrained deep learning model

Inspired by pre-training and fine-tuning techniques of large models, researchers have developed an advanced convolutional neural network framework specifically for accurately reconstructing data from X-ray scanning coherent diffraction imaging experiments, improving phase recovery precision and efficiency.

Figure 2: Efficient reconstruction of phase information in Ptychography using a fine-tuning strategy for large models

Pan, Xinyu et al. “An efficient ptychography reconstruction strategy through finetuning of large pre-trained deep learning model.” iScience 26 (2023):108420.

2. A novel method for 6D WAXD tensor tomography based on "Virtual Reciprocal Space Scanning"

To address the complexity and low timeliness of traditional physical analysis methods for high-throughput multidimensional diffraction experiments, a novel 6D tensor tomography method based on "virtual reciprocal space scanning" is proposed. This reduces the experimental scanning dimensions from four to three. Using a deep fully connected neural network, the hidden 3D fiber orientation distribution information in massive 2D diffraction images is extracted, enabling high-precision online analysis of diffraction data and achieving a 10,000-fold increase in analytical efficiency.

Figure 3: A new method for 6D tensor tomography imaging based on "virtual inverted space scanning"

Zhao, Xiaoyi et al. “A step towards 6D WAXD tensor tomography.” IUCrJ 11 (2024): 502 - 509.

3. Systematic Diffraction Image Denoising Algorithm for In-Situ Dynamic Experiments

For in-situ dynamic experiments, a systematic denoising tool has been developed, including supervised SEDCNN and unsupervised SMAE algorithms, which significantly outperform existing state-of-the-art (SOTA) models in training speed and denoising performance, effectively enhancing the recovery of physical information in in-situ dynamic experiments.

Figure 4-1: Supervised image denoising algorithm for in-situ dynamic experiments - SEDCNN

Figure 4-2: Unsupervised image denoising algorithm for in-situ dynamic experiments - SMAE

Zhou, Zhongzheng et al. “A machine learning model for textured X-ray scattering and diffraction image denoising.” npj Computational Materials 9 (2023): 1-14.

4. End-to-End Image Jitter and Offset Correction Method for Synchrotron Imaging Experiments

To address the demand for image jitter correction in high spatial resolution synchrotron imaging experiments, an end-to-end image correction method combining sample contour structures and AI-based image enhancement techniques is proposed. This method has achieved excellent performance in image correction tasks for various nano-precision synchrotron imaging methods.

Figure 5: End to end synchrotron radiation imaging experimental image mixing jitter and offset correction method

Zhang, Zhen et al. “A general image misalignment correction method for tomography experiments.” iScience 26 (2023): 107932.

5. Rapid Analysis of Fine Structural Changes in Angle-Resolved Photoelectron Spectroscopy Data Using Higher-Order Fine Clustering Algorithms

By utilizing higher-order fine clustering algorithms, researchers can quickly analyze fine structural changes in scanning angle-resolved photoelectron spectroscopy (NanoARPES) data, potentially revolutionizing ARPES experimental workflows and improving experimental accuracy and efficiency.

Figure 6: High order fine clustering algorithm for analyzing fine structure changes in angular resolved photoelectron spectroscopy experimental data

6. Dynamically Adaptable Synchrotron Experiment Image Deconvolution and Super-Resolution Pipeline

Based on Fourier attention mechanism-driven generative adversarial networks (GANs), researchers have proposed a dynamically adaptive algorithm for synchrotron image super-resolution workflow. By organically integrating and deploying cuttingedge algorithms online, this workflow forms a complete closed loop of experimental decision-making, data acquisition, online processing, real-time analysis, feedback control, and experimental decision-making.

Figure 7: Algorithm dynamically adaptive deconvolution and super-resolution workflow

Chun Li,Xiaoxue Bi,Yujun Zhang,Zhen Zhang,Liwen Wang,Jian Zhuang,Dongliang Chen,Yuhui Dong,Yi Zhang.Towards adaptable synchrotron image restoration pipeline.

7. Towards full-stack deep learning-empowered data processing pipeline for synchrotron tomography experiments

The research team has introduced a deep learning-enabled full-stack data processing pipeline for synchrotron tomography experiments. They highlight that large models for synchrotron tomography, intelligent scheduling centers, and long-term learning strategies at beamlines will become effective solutions to the big data challenges faced in tomography experiments.

Figure 8: Data processing pipeline for full stack synchrotron radiation tomography experiment empowered by deep learning

Zhang, Zhen et al. “Towards full-stack deep learning-empowered data processing pipeline for synchrotron tomography experiments.” The Innovation 5 (2023):100539.