串行晶体学数据筛选算法研究

Research on serial crystallography data screening algorithm

在串行飞秒晶体学中,可以使用多种基于机器学习的方法对数据进行筛选分类;在晶体实验中,采用自动化图像处理和卷积神经网络来检测晶体衍射图中的布拉格点,去除无效的实验数据。对不同的样品采用多种机器学习方法,进行多次实验和模拟,分析实验预测结果准确度不同的原因。结果显示:在众多方法中,卷积神经网络能够得到较高的预测准确率,而线性方法仅对某些样品有较好的准确率,但两者都优于传统找点算法。为X射线自由电子激光(X-ray Free Electron Laser,XFEL)实验提供有效和便捷的数据筛选工具。

[Background] A number of machine learning-based methods have been developed for classification screening of serial femtosecond crystallography (SFX) data.[Purpose] This study aims to improve the data analysis technique to obtain more effectively screen SFX data from X-ray free electron lasers (XFELS).[Methods] First of all, experimental data of different samples were taken for screening and categorizing, Automatic image processing and convolutional neural network were employed for detection of Bragg points in diffraction patterns of crystals, hence, invalid experimental data were filtered out. Then, simulation was performed as the benchmark for many experiments. Finally, the reasons for the different accuracy of experimental prediction results were analyzed.[Results] Among many methods, convolutional neural network can obtain higher prediction accuracy whilst linear method only has better accuracy for some samples, but both of them are superior to traditional point-finding algorithm.[Conclusions] This approach provides an effective and convenient data screening tool for X-ray FEL experiments.