The TrackML high-energy physics tracking challenge on Kaggle

Abstract

The High-Luminosity LHC (HL-LHC) is expected to reach unprecedented collision intensities, which in turn will greatly increase the complexity of tracking within the event reconstruction. To reach out to computer science specialists, a tracking machine learning challenge (TrackML) was set up on Kaggle by a team of ATLAS, CMS, and LHCb physicists tracking experts and computer scientists building on the experience of the successful Higgs Machine Learning challenge in 2014. A training dataset based on a simulation of a generic HL-LHC experiment tracker has been created, listing for each event the measured 3D points, and the list of 3D points associated to a true track.The participants to the challenge should find the tracks in the test dataset, which means building the list of 3D points belonging to each track.The emphasis is to expose innovative approaches, rather than hyper-optimising known approaches. A metric reflecting the accuracy of a model at finding the proper associations that matter most to physics analysis will allow to select good candidates to augment or replace existing algorithms.

A Common Tracking Software Project

“…An early version of ACTS has been used to simulate the dataset for the Tracking Machine Learning (TrackML) challenge [24][25][26], which was performed in two stages to invite collaborators from within and external to particle physics to stimulate the development of new ideas for track reconstruction. The dataset produced for this challenge has subsequently been used to explore a range of novel track reconstruction algorithms [12,[27][28][29][30].…”

Towards a realistic track reconstruction algorithm based on graph neural networks for the HL-LHC

“…The Generic Detector has also been used to simulate the dataset for the TrackML Challenge [9]. It is shown in Figure 1.…”

“…Graphs created with this cut comprise O(10 4 ) nodes and O(10 5 ) edges. Using this sample, the training of the model described above fits into the memory of the Quadro RTX 8000 and also into the 32 GB of memory of an Nvidia Tesla V100 GPU 9 , and no swapping is needed. Overfitting is seen after 60000 iterations, which take 12 hours to be completed.…”

“…This includes improvements of existing algorithms and the development of new algorithms, possibly incorporating machine learning (ML) for track pattern recognition [8]. To boost the development of ML techniques for tracking and strengthen the involvement of ML experts in this effort, a challenge on the Kaggle platform has been organised in 2018-2019: the TrackML Challenge [9]. At the end of the challenge, ML was not at the core of the fastest algorithms [10], and in general the proposed solutions included less innovative ML than expected.…”

Towards a realistic track reconstruction algorithm based on graph neural networks for the HL-LHC

“…This includes improvements of existing algorithms and the development of new algorithms, possibly incorporating machine learning (ML) for track pattern recognition [8]. To boost the development of ML techniques for tracking and strengthen the involvement of ML experts in this effort, a challenge on the Kaggle platform has been organised in 2018-2019: the TrackML Challenge [9]. At the end of the challenge, ML was not at the core of the fastest algorithms [10], and in general the proposed solutions included less innovative ML than expected.…”