Shanghai Optics and Machinery Institute has made progress in near-field state analysis of high-power laser devices based on convolutional neural networks

Recently, the research team of the High Power Laser Physics Joint Laboratory of the Chinese Academy of Sciences Shanghai Institute of Optics and Fine Mechanics identified and analyzed the abnormal near-field output of the SG - Ⅱ upgrade device by using the spatial domain computing method and the deep learning model with attention mechanism in response to the requirements of real-time and effectiveness of the analysis of the multi-channel near-field output of the high power laser device. The relevant results were published in Optics and Lasers in Engineering under the title "Near field analysis of the high-power laser facility using calculated methods and a residual convolutional neural network with attention mechanism".

The physical research of Inertial Confinement Fusion (ICF) has put forward very strict requirements for the output performance and reliability of high-power laser drivers. Among them, the uniform distribution of the near-field is conducive to improving the operating flux of the system, protecting subsequent optical components, and meeting the requirements of long-term high-intensity reliable operation of the system. High power laser devices contain multiple beams of laser, and manual identification methods are not timely and effective enough. Therefore, effective methods are needed to analyze the near-field state at different times and provide timely warnings. Convolutional neural networks (CNNs) have powerful feature extraction capabilities and can be trained on historical data to meet complex and diverse task requirements.

Researchers have proposed using spatial computing methods and residual convolutional neural network models with additional attention mechanisms to preliminarily evaluate the operation of the SG-II upgrade device based on a large number of near-field images at different times. The airspace calculation method is used for batch processing of near-field images detected by CCD, and can analyze the changes in near-field distribution uniformity during the continuous operation time of the device through adjustments and contrast analysis. This algorithm automatically extracts effective near-field spot regions, which also provides a preprocessing step for images used to train convolutional neural network models. 
Convolutional neural network models are used for automatic recognition and classification of near-field image features with multiple labels, thereby achieving fundamental frequency (1) ω）  Detection of near-field state anomalies. In this work, researchers selected six features including near-field distribution uniformity, abnormal output signals, and strong diffraction rings for analysis. The classification accuracy of the model reached 93%, and the model can make real-time judgments about the above six features for any number of near-field images.

In subsequent research, as the amount of experimental data increases, researchers will refine the classification labels for abnormal features, especially similar features, to establish more robust models. This work explores the effective application of deep learning models in high-power ICF laser devices, and is expected to continue expanding the application scope of deep learning models in the future, providing intelligent analysis methods for large-scale laser devices.

Figure 1: Results of Spatial Calculation Method (a) CCD Capture Image (b) Near Field Grayscale Distribution Histogram (c) Near Field Grayscale Distribution Histogram After Removing Background (d) Binary Image After Removing Background (e) Rotated Near Field Image After Hough Transform (f) Rotated Binary Image (g) Cropped Near Field Image (h) 85% Region of Near Field Image

Figure 2 Structure of Spatial Attention Residual Convolutional Neural Network Model

Source: Shanghai Institute of Optics and Fine Mechanics, Chinese Academy of Sciences