Scholarlink

Purpose <p>Optimizing synchrotron radiation beamlines is critical prior to conducting experiments, but typically requires several hours or even days when performed manually. Although intelligent optimization methods have brought about improvements, the beam time required for their online testing remains substantial, given that beam time is limited. This study leverages the test beamline (BL10U) at the Hefei Advanced Light Facility (HALF) as the physical entity to address the challenges.</p> Methods <p>In the study, we developed a virtual beamline platform based on the experimental physics and industrial control system (EPICS). A simulation model of BL10U was constructed using XRayTracer (XRT) software to generate simulation data. To fully exploit the data, an artificial neural network (ANN) surrogate model was trained via <i>K</i>-fold cross-validation. Additionally, based on the virtual beamline platform, this study employed the differential evolution (DE), non-dominated sorting genetic algorithm Ⅱ (NSGA-Ⅱ), and particle swarm optimization (PSO) algorithms to conduct offline beamline tuning practices and optimize multiple optical element (OE) parameters using both the XRT simulation and ANN surrogate model.</p> Results <p>The ANN surrogate model achieved a prediction accuracy of 97% on the test dataset and performed approximately 3000 times faster than the XRT model. When coupled with the DE, NSGA-Ⅱ and PSO algorithms, the ANN model converged in 65.0, 5.1 and 9.3&#xa0;s, respectively, compared to 16.5, 5.0, and 3.8&#xa0;h for XRT simulation. Among these, the optimization based on DE algorithm achieved the best beam performance.</p> Conclusion <p>The proposed virtual beamline platform provides a safe and efficient offline environment supporting intelligent optimization methods testing. This approach marks a significant step toward transitioning from manual to intelligent beamline adjustment, enhancing operational efficiency. Experiments with several algorithms on the platform show that multiple algorithms can work effectively on this platform, and the results have good consistency.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Research of optimization methods on the ANN-based virtual beamline platform

  • 期刊论文

  • 发表日期：

    2025年12月19日

查看全文
  • Jiahui Zhang,
  • Xueting Wu,
  • Lingfei Hu,
  • Anbo Kong,
  • Bo He,
  • Liuguo Chen,
  • Gongfa Liu

摘要

Purpose

Optimizing synchrotron radiation beamlines is critical prior to conducting experiments, but typically requires several hours or even days when performed manually. Although intelligent optimization methods have brought about improvements, the beam time required for their online testing remains substantial, given that beam time is limited. This study leverages the test beamline (BL10U) at the Hefei Advanced Light Facility (HALF) as the physical entity to address the challenges.

Methods

In the study, we developed a virtual beamline platform based on the experimental physics and industrial control system (EPICS). A simulation model of BL10U was constructed using XRayTracer (XRT) software to generate simulation data. To fully exploit the data, an artificial neural network (ANN) surrogate model was trained via K-fold cross-validation. Additionally, based on the virtual beamline platform, this study employed the differential evolution (DE), non-dominated sorting genetic algorithm Ⅱ (NSGA-Ⅱ), and particle swarm optimization (PSO) algorithms to conduct offline beamline tuning practices and optimize multiple optical element (OE) parameters using both the XRT simulation and ANN surrogate model.

Results

The ANN surrogate model achieved a prediction accuracy of 97% on the test dataset and performed approximately 3000 times faster than the XRT model. When coupled with the DE, NSGA-Ⅱ and PSO algorithms, the ANN model converged in 65.0, 5.1 and 9.3 s, respectively, compared to 16.5, 5.0, and 3.8 h for XRT simulation. Among these, the optimization based on DE algorithm achieved the best beam performance.

Conclusion

The proposed virtual beamline platform provides a safe and efficient offline environment supporting intelligent optimization methods testing. This approach marks a significant step toward transitioning from manual to intelligent beamline adjustment, enhancing operational efficiency. Experiments with several algorithms on the platform show that multiple algorithms can work effectively on this platform, and the results have good consistency.