Graph-based deep learning for representing events at the Large Hadron Collider

找到合适的特定领域的数据表示法是解决科学问题的关键。当代机器学习已经开发出了一些体系结构，这些体系结构在某些领域被证明是成功的，提出了捕捉大数据集中相关性的自适应隐藏变量。欧洲核子研究中心的大型强子对撞机(LHC)是一个非常大的数据集的来源，这些数据集提供了学习新的基础物理的可能性，因此需要开发新的模式识别任务。从已知物理到大型强子对撞机探测器中的观测结果的映射构成了寻求新颖性的背景。这种映射需要将物理约束转化为机器学习框架，以便校准数据流并提出优化目标，以便允许现代深度学习方法提供潜在的解决方案。该项目中将探索的一种有前景的方法是使用图形卷积网络将检测到的粒子阵雨事件嵌入为点云。这将提供对撞机事件的表示，可用于处理任务，例如从背景中检测适当的新物理信号。该项目将与南安普顿大学和卢瑟福·阿普顿实验室的粒子物理学家共同监督，他们与大型强子对撞机有现有的联系。粒子物理方面的适当培训将作为项目开始时的一部分，通过下一届博士生和研究生讲座提供。

Finding appropriate domain specific representations of data is key to solving problems in science. Contemporary machine learning has developed architectures that have proved to be successful in some domains in proposing adaptive hidden variables that capture correlations in large datasets. The Large Hadron Collider (LHC) in CERN is a source of very large datasets that offer the possibility to learn new fundamental physics for which novel pattern recognition tasks need to be developed. The mapping from the known physics to the observations in the detectors of the LHC constitutes the background against which novelties are sought. This mapping requires the necessary translation of physical constraints into machine learning frameworks in order to calibrate data streams and propose optimisation objectives in order to allow modern deep learning methods to offer potential solutions. One of the promising approaches to be explored in this project is the use of graph convolution networks to embed events of detected particle showers treated as point clouds. This will provide the representations of collider events that can be used for processing tasks such as detecting appropriate signals of new physics from background. This project will be jointly supervised with particle physicists from the University of Southampton and Rutherford Appleton Laboratory who have existing links with LHC. Appropriate training in particle physics will be provided as part of the studentship at the start of the project via the NExT PhD School and the RAL Graduate Lectures.