An international team launches a great machine learning contest, TrackML. The goal: to link myriads of points to trace trajectories and discover new particles.
A contest as a research tool
In the field of machine learning, competitions are a very good way to evaluate the performance of algorithms and highlight the most effective ones. This also allows contest organizers to put a particular problem under huge spotlights. Used in economics, climate science, astrophysics, chemistry or high-energy physics, competitions have gradually become genuine research tools in their own right.
The TrackML competition, organized by an international team, takes place in the context of particle physics since it aims to improve the detection and characterization of new particles created within the large hadron collider (LHC). The LHC is located at CERN and is the world's largest particle accelerator. High-energy physics researchers use this accelerator to collide very high-energy protons, then carefully scan the "waste" from the collision for new particles.
Part of the tunnel containing the LHC, the largest and most powerful particle accelerator in the world (Image credit: CERN)
Connecting dots to discover new particles
Each collision (also called "event") creates a firework of particles in all directions, which pass through detectors leaving some of their energy there. The aim is to be able to trace particle trajectories only from the impact points on the detectors: i.e. assign each measurement point to a particular particle.
Indeed, each type of particle has its own trajectory, and it is therefore possible to identify which ones have been created by tracing all the trajectories and comparing them with those of already known particles. The algorithms which will compete will have to find these trajectories starting from a list of points in three dimensions.
Starting with 10,000 3D maps each containing 10,000 impact points, these algorithms will have to be able to count the number of particles created at each event, attribute their impacts (about 12 impacts per particle), and trace the trajectories of all these particles. The perfect algorithm would be capable of perfectly associating each impact on a detector with the trajectory from which it originated.
Examples of cross-section views of simulated 3D maps of sensor impacts and trajectories to be reconstructed.
Increase calculation accuracy and speed
There are already many methods and algorithms for this task, but the goal of the contest organizers is to get a more accurate method than anything that is done today, while significantly increasing the speed of the calculations. These improvements are necessary because an increase in LHC performance is expected in 2025. More detectors, and more sensitive detectors, which will increase the number of impacts at each event to 10000 from 1000 today. More data will be produced that will need to be processed ten times faster.
The raw data from which the algorithms will attempt to plot the trajectories are derived from simulations of which only a handful of organizers know the real trajectories. This will limit the risk of cheating, and they will be able to objectively evaluate which algorithms are best.
The winning team will receive €25,000 and will be invited to present their method at CERN, so don't wait to start!