Collaborative Research: SI2:SSE: Extending the Physics Reach of LHCb in Run 3 Using Machine Learning in the Real-Time Data Ingestion and Reduction System

在过去的200年里，物理学家发现了普通物质的基本成分，并发展了一种非常成功的理论来描述它们之间的相互作用(力)。所有的原子和构成原子的分子都可以用这些成分来描述。原子核通过强烈的核相互作用结合在一起。它们的衰变是由强核相互作用和弱核相互作用造成的。电磁力将原子结合在一起，并将原子结合成分子。用量子场论描述了电磁力、弱核力和强核力。这些理论的预测可以非常非常精确，而且它们已经得到了同样精确的实验测量的验证。最近，位于瑞士欧洲核子研究中心实验室的大型强子对撞机(LHC)发现了一种新的基本粒子，希格斯玻色子，这是统一弱相互作用和电磁相互作用所必需的。尽管在过去的一个世纪里获得了关于自然基本粒子和力的大量知识，但许多重要的问题仍然没有得到回答。例如，宇宙中大多数通过引力相互作用的物质并不存在普通的电磁或核相互作用。由于它只能通过引力相互作用被观察到，所以它被称为暗物质。那是什么？同样有趣的是，当我们所知道的基本相互作用将物质和反物质描述为彼此近乎完美的镜像时，为什么宇宙中几乎没有反物质？建造大型强子对撞机是为了发现和研究希格斯玻色子，并寻找这些问题的答案。大型强子对撞机的第一次数据采集(2010年至2012年)取得了巨大的成功，发表了1000多篇期刊论文，希格斯玻色子的发现凸显了这一点。目前的大型强子对撞机运行(第二次运行，2015年至今)已经产生了许多世界领先的结果；然而，最有趣的问题仍然没有得到回答。位于欧洲核子研究中心大型强子对撞机上的LHCb实验具有独特的潜力来回答其中一些问题。LHCb正在大型强子对撞机上搜索高能粒子碰撞产生的暗物质信号，并对罕见的过程进行高精度研究，这些过程可能揭示导致我们宇宙中观察到的物质/反物质失衡的未知力量的存在。该项目的主要目标--由计算机和信息科学与工程局的高级数字基础设施办公室和数学和物理科学局的物理司和数学科学司支持--是开发和部署利用机器学习(ML)的软件，使LHCb实验能够在第3轮(2021-2023年)显著提高其发现潜力。具体地说，开发的ML将通过使用可用的有限计算资源，使更有效地识别和研究潜在信号成为可能，从而极大地提高对许多拟议类型的暗物质和新作用力的敏感度。大型强子对撞机实验收集的数据集是世界上最大的。例如，LHCb实验的传感器阵列，两个PI都在工作，每秒产生大约100太字节的数据，接近每年1ZB的数据。即使在定制读出电子设备执行了大幅数据缩减后，数据量仍约为每年10艾字节，可与最大规模的工业数据集相媲美。如此庞大的数据集不可能被无限期地存储；因此，所有高能物理(HEP)实验都采用了由数据摄取系统实时执行的数据简化方案-在HEP中称为触发系统-来决定是将每个事件保留下来供未来分析，还是永久丢弃。触发系统的设计取决于传感器的读出速度、数据摄取系统的计算能力以及数据的可用存储空间。LHCb探测器正在为第3次(2021-2023)进行升级，届时触发系统将需要每年处理25艾字节。目前，触发器每年处理的10艾字节中只有0.3使用高级计算算法进行分析；其余的在此阶段之前使用在现场可编程门阵列上执行的简单算法丢弃。为了处理CPU场中的所有数据，将使用ML来开发和部署新的触发算法。该建议的具体目标是使用ML更全面地表征LHCb数据，并使用这些表征来构建算法：替换事件模式识别中计算代价最高的部分；提高事件分类算法的性能；并在不降低物理性能的情况下减少每个事件持续存在的字节数。由于触发系统的限制，暗物质和宇宙的物质/反物质不对称性的许多可能的解释目前还无法得到。由于未来HEP计算预算预计将大致持平，LHCb触发系统必须重新设计以实现其全部潜力。这种重新设计必须超越可扩展的技术升级；需要激进的新战略。

In the past 200 years, physicists have discovered the basic constituents of ordinary matter and the developed a very successful theory to describe the interactions (forces) between them. All atoms, and the molecules from which they are built, can be described in terms of these constituents. The nuclei of atoms are bound together by strong nuclear interactions. Their decays result from strong and weak nuclear interactions. Electromagnetic forces bind atoms together, and bind atoms into molecules. The electromagnetic, weak nuclear, and strong nuclear forces are described in terms of quantum field theories. The predictions of these theories can be very, very precise, and they have been validated with equally precise experimental measurements. Most recently, a new fundamental particle required to unify the weak and electromagnetic interactions, the Higgs boson, was discovered at the Large Hadron Collider (LHC), located at the CERN laboratory in Switzerland. Despite the vast amount of knowledge acquired over the past century about the fundamental particles and forces of nature, many important questions still remain unanswered. For example, most of the matter in the universe that interacts gravitationally does not have ordinary electromagnetic or nuclear interactions. As it has only been observed via its gravitation interactions, it is called dark matter. What is it? Equally interesting, why is there so little anti-matter in the universe when the fundamental interactions we know describe matter and anti-matter as almost perfect mirror images of each other? The LHC was built to discover and study the Higgs boson and to search for answers to these questions. The first data-taking run (Run 1, 2010-2012) of the LHC was a huge success, producing over 1000 journal articles, highlighted by the discovery of the Higgs boson. The current LHC run (Run 2, 2015-present) has already produced many world-leading results; however, the most interesting questions remained unanswered. The LHCb experiment, located on the LHC at CERN, has unique potential to answer some of these questions. LHCb is searching for signals of dark matter produced in high-energy particle collisions at the LHC, and performing high-precision studies of rare processes that could reveal the existence of the as-yet-unknown forces that caused the matter/anti-matter imbalance observed in our universe. The primary goal of this project - supported by the Office of Advanced Cyberinfrastructure in the Directorate for Computer and Information Science and Engineering and the Physics Division and the Division of Mathematical Sciences in the Directorate of Mathematical and Physical Sciences - is developing and deploying software utilizing Machine Learning (ML) that will enable the LHCb experiment to significantly improve its discovery potential in Run 3 (2021-2023). Specifically, the ML developed will greatly increase the sensitivity to many proposed types of dark matter and new forces by making it possible to much more efficiently identify and study potential signals -- using the finite computing resources available. The data sets collected by the LHC experiments are some of the largest in the world. For example, the sensor arrays of the LHCb experiment, on which both PIs work, produce about 100 terabytes of data per second, close to a zettabyte of data per year. Even after drastic data-reduction performed by custom-built read-out electronics, the data volume is still about 10 exabytes per year, comparable to the largest-scale industrial data sets. Such large data sets cannot be stored indefinitely; therefore, all high energy physics (HEP) experiments employ a data-reduction scheme executed in real time by a data-ingestion system - referred to as a trigger system in HEP - to decide whether each event is to be persisted for future analysis or permanently discarded. Trigger-system designs are dictated by the rate at which the sensors can be read out, the computational power of the data-ingestion system, and the available storage space for the data. The LHCb detector is being upgraded for Run 3 (2021-2023), when the trigger system will need to process 25 exabytes per year. Currently, only 0.3 of the 10 exabytes per year processed by the trigger are analyzed using high-level computing algorithms; the rest is discarded prior to this stage using simple algorithms executed on FPGAs. To process all the data on CPU farms, ML will be used to develop and deploy new trigger algorithms. The specific objectives of this proposal are to more fully characterize LHCb data using ML and build algorithms using these characterizations: to replace the most computationally expensive parts of the event pattern recognition; to increase the performance of the event-classification algorithms; and to reduce the number of bytes persisted per event without degrading physics performance. Many potential explanations for dark matter and the matter/anti-matter asymmetry of our universe are currently inaccessible due to trigger-system limitations. As HEP computing budgets are projected to be approximately flat moving forward, the LHCb trigger system must be redesigned for the experiment to realize its full potential. This redesign must go beyond scalable technical upgrades; radical new strategies are needed.