A team from the Chinese Academy of Sciences uses AI large model training technology to process massive synchrotron radiation data

Without changing the original meaning, the sentence that needs to be rewritten into Chinese is: Edit| It is a coherent diffraction imaging technology that can theoretically achieve diffraction-limited resolution and has been widely used in research in various scientific fields such as materials, life, semiconductors, and energy.

The new generation of synchrotron radiation light source can provide high coherence and high brightness X-rays, promoting the development of coherent imaging technology in the direction of high-throughput and multi-dimensional, making ptychography useful in fine structure research and functional characterization of large-volume samples It has excellent application prospects. However, new experimental models and application scenarios have brought technical challenges to the online analysis of massive data. The amount of original diffraction pattern data for a single experiment can reach the PB level, becoming one of the largest data sources for scientific experiments on the fourth-generation synchrotron radiation light source. one. In addition, its phase recovery problem is also one of the most difficult problems in the field of synchrotron radiation data processing.

As a powerful tool for big data analysis and processing, artificial intelligence methods maintain the advantages of traditional algorithms and highlight their capabilities in online analysis of massive experimental data.

As a relatively time-consuming scanning imaging technology, one of the main goals of ptychography is to enable real-time analysis. However, it is difficult for the current traditional ptychography reconstruction algorithm to meet the needs of online reconstruction. Based on the convolutional neural network, the research team proposed a grouped convolutional neural network decoder structure, which makes the network training and reconstruction faster and the reconstruction effect better. Neural networks can learn to map from diffraction patterns to real objects. Thanks to the further improvement in the volume and quality of light source data in the future, the network scale, parameter volume, and training data volume will further increase, which will improve the performance and generalization capabilities of the network. The High Energy Synchrotron Radiation Source (HEPS) beamline software team of the Chinese Academy of Sciences has developed a convolutional neural network framework called PtyNet to recover objects from X-ray Ptychography experimental data accurate projection. With the support of powerful computing clusters, PtyNet can quickly obtain data from synchrotron radiation light sources for training, and quickly reconstruct images of users’ experimental data

Figure 1

The research is titled "Efficient ptychography reconstruction strategy by fine-tuning large pre-trained deep learning models" and was published in iScience magazine on November 9, 2023

Paper link:

https://doi.org/10.1016/j.isci.2023.108420

Since the target objects recovered by different experimental data are different, the team also introduced a fine-tuning strategy Further optimize network parameters. The unsupervised fine-tuning strategy enables the network to have stronger generalization ability and higher reconstruction resolution. Synchrotron radiation sources can provide the network with sufficient data to obtain a more powerful pre-trained model. Even for a new sample that does not appear within the network, the network can be successfully reconstructed (Figure 2).

What needs to be rewritten is: The second picture

In the future, the team will continue to apply convolutional neural networks to X-ray coherence imaging research in the field. Using fine-tuning and large model strategies, a large model of coherent imaging was developed. The model itself can identify different imaging tasks and give recovery results. Users only need to input a few line station parameters for real-time reconstruction.

Facing the challenge of EB-scale data in the future, HEPS is actively promoting the innovative scientific research paradigm of "Large-scale scientific software framework AI for Science" and has established a professional scientific software team to carry out experimental control, large-scale Cross-field research such as data collection and processing, artificial intelligence, cutting-edge subject algorithms, multi-scale image processing and data mining has laid the foundation for the construction of "smart light sources".

The above is the detailed content of A team from the Chinese Academy of Sciences uses AI large model training technology to process massive synchrotron radiation data. For more information, please follow other related articles on the PHP Chinese website!