Big Data Analysis Techniques Applied to the NA62 Experiment at CERN

欧洲核子研究中心的NA62实验旨在精确测量一个正电介子衰变为一个正电介子和两个中微子（Kpnn）的10-10分支比。2016年，该实验收集了足够的k介子衰变来观察Kpnn；2017年，NA62的数据增加了10倍，预计2018年将有可比的统计数据。对2016年数据的分析主要使用了一种基于切割的技术，适用于原始水平的数量。这是一种经过验证的观察Kpnn的方法，但它不能为精确的分支比测量提供足够的信号接受度，而且它不容易扩展到更高的统计数据样本。本项目旨在为NA62开发一种有效的数据缩减方案，并将多变量技术应用于Kpnn分析。2017年，NA62已经产生了超过1拍字节的原始数据，这些数据正在使用传统的高能物理（HEP）分析模型进行处理。这种模型采用了几个阶段的数据处理：校准，重建，数据质量评估，过滤。物理分析只能从过滤的数据集开始，因为计算吞吐量的考虑，访问数据质量和校准信息是至关重要的。从数据科学的角度来看，这种方法有几个值得注意的特点。首先，从物理学家的角度来看，中间重建数据集（预计到2018年底将达到10拍字节）是无用的。尽管如此，数据科学方法仍然需要对输入的原始数据进行注释，以生成经过校准的原始数据集，然后可以在分析级别进行查询。其次，重建数据集的大小是输入原始数据集的三倍：这对于实验的调试阶段是正常的，但不适用于开发阶段。拟议的数据科学项目是为了解决NA62分析模型目前的局限性，并减少产生物理结果的时间。在项目的第一阶段，将研究重构的数据大小和I/O性能，以期利用NA62计算资源减小数据大小并提高I/O吞吐量。重建和过滤的数据集都需要进行研究，以确保最终用户和批量数据处理工作流程的最佳性能。在第二阶段，分析模型本身将被研究，以潜在地利用数据科学方法来减少物理洞察的时间。有几种方法值得研究，包括数据同质化（减少分析代码的复杂性）和分析同质化（减少分析工作流的复杂性）的可能性。分析模型改进的新方法包括火花式分析，它需要一个专门的分析工具来提供基础设施，当考虑将来采用更统一的方法来支持HEP计算需求时，这种基础设施可能会非常有趣。作为新分析模型有效性的最终测试，该项目旨在将机器学习技术应用于Kpnn分析，目的是提高信号接受度。这些技术的发展将利用数据简化来有效地创建和优化训练、验证和测试样本，这是任何成功的机器学习应用于数据分析的核心。将研究这些技术对粒子识别、光子抑制和跟踪的影响，并研究几种算法，使用特定的HEP包（如TMVA），但也探索HEP之外的解决方案，如scikit-learn或Keras包。

The NA62 experiment at CERN aims to measure precisely the 10-10 branching ratio of the decay of a positive kaon into a positive pion and two neutrinos (Kpnn). In 2016 the experiment has collected enough kaon decays to observe Kpnn; in 2017 NA62 has taken a factor 10 more data and a comparable statistics is expected in 2018. The analysis of the 2016 data has been carried on using mostly a cut-based technique applied to raw-level quantities. This is a proven method to observe Kpnn, but it does not provide enough signal acceptance for a precise branching ratio measurement and it is not easily scalable to higher statistics data samples. The present project aims to develop an efficient data reduction scheme for NA62 and to apply multi-variate techniques to the Kpnn analysis. Already in 2017 NA62 produced over a petabyte of raw data which are under processing using a traditional high energy physics (HEP) analysis model. Such a model employs several stages of data processing: calibration, reconstruction, data quality assessments, filtering. Physics analysis can only starts on filtered datasets because of computing throughput considerations, with access to data quality and calibration information being crucial. From the data science perspective there are several noteworthy features of this approach. The first is that the intermediate reconstructed dataset, expected of order of 10 petabytes at the end of 2018, is useless from the physicist's perspective. Nevertheless, a data science approach would still involve annotating the input raw data to produce a calibrated raw dataset which could then be queried at analysis level. Secondly the reconstructed dataset is three times the size of the input raw dataset: this is normal for the commissioning phase of an experiment, but not applicable in exploitation phase. The proposed data science project is to address the current limitations of the NA62 analysis model and reduce the time to produce physics results. In the first stage of the project, the reconstructed data size and I/O performance will be studied with a view to reducing the size and improving the I/O throughput using NA62 computing resources. Both the reconstructed and the filtered datasets will need to be studied to ensure the best performance for the end-user and for bulk data processing workflows. In the second stage, the analysis model itself will be studied with a view to potentially leveraging data science approaches to reduce the time to physics insight. Several approaches are worthy of investigation, including the possibilities for data homogenisation (reducing the complexity of analysis code) and analysis homogenisation (reducing the complexity of analysis workflows). Novel approaches to analysis model improvements include spark-style analysis, requiring a dedicated analysis facility to provide the infrastructure that could be extremely interesting when considering a more unified approach to supporting HEP computing demands in future.As a final test of the effectiveness of the new analysis model, the project aims to apply machine learning techniques to the Kpnn analysis with the goal to increase the signal acceptance. The development of these techniques will take advantage from data reduction to efficiently create and optimize training, validation and testing samples, which are the core of any successful machine learning application to data analysis. The impact of these techniques on particle identification, photon rejection and tracking will be studied and several algorithm investigated, using specific HEP-packages like TMVA, but also exploring solutions outside HEP, like scikit-learn or Keras packages.