Graph Neural Networks for pattern recognition and Fast Track Finding in LHC data

在粒子物理对撞机实验中，在探测器中重建粒子轨迹(轨迹)是实验数据分析中最具挑战性的计算和智力部分之一。一个典型的大型强子对撞机探测器包含数千个传感器，测量粒子沿其轨迹的位置，轨迹发现(又名模式识别)问题是将单独的测量(“命中”)关联到代表粒子轨迹的序列。这个问题的规模是巨大的，因为每个活动的点击量可能高达400000；曲目的数量--数千。目前大型强子对撞机实验中所采用的寻迹算法是基于组合航迹跟踪算法，该算法随着命中次数的增加而呈非线性扩展。相应的CPU时间增加，通常接近立方，产生了对计算能力的巨大且不断增长的需求。这一缺陷促使人们研究新的轨迹发现方法，特别是基于机器学习(ML)技术的方法。这种方法的好处可以节省数千万英镑的CPU。本项目的目的是探索利用基于图的航迹模型和图神经网络(GNN)的航迹查找方法。将开发一种基于ML的算法，以迭代地解决图网络中的歧义连接，并在给定输入命中特征(例如，表示硅探测器中的轨道位置测量的电荷簇的形状)的情况下提取良好的候选轨道。每个GNN神经元的兴奋/抑制规则将被设计为促进“命中到轨迹”关联的“简单到复杂”的方法，从而使网络从具有低命中密度的事件的相对“容易”区域开始，并逐渐向更复杂的“热点”区域发展。

In particle physics collider experiments, reconstructing particle trajectories (tracks) in the detector is computationally and intellectually one of the most challenging parts of experimental data analysis. A typical LHC detector contains many thousands of sensors measuring particle positions along their trajectory and the track finding (aka pattern recognition) problem is to associate individual measurements ("hits") into sequences representing particle trajectories. The scale of the problem is enormous, given that the number of hits can be up to 400000 per event; the number of tracks -- several thousands. The current track finding algorithm adopted in the LHC experiments is based on the combinatorial track following which scales non-linearly with the number of hits. The corresponding CPU time increase, typically close to cubical, creates huge and ever-increasing demand for computing power. This drawback motivates investigating novel approaches for track finding, in particular, those based on the machine learning (ML) techniques. The benefits of such an approach could lead to tens of millions of pounds savings in CPU. The aim of this project is to explore track finding methods utilizing a graph-based track model and graph neural networks (GNN). An ML-based algorithm will be developed to iteratively resolve ambiguous connections in the graph network and extract good track candidates, given the input hit features such as a shape of a charge cluster representing a track position measurement in a silicon detector. The excitation/inhibition rules of individual GNN neurons will be designed to facilitate the "simple-to-complex" approach for "hits-to-tracks" association such that the network starts with relatively "easy" areas of an event with low hit density and gradually progresses towards more complex "hot" areas.