Understanding Particle Physics: Data Science Jets into the Future

Bruna, Cho, and Cranmer apply new neural network to jet classification problem

The Large Hadron Collider, the most powerful particle collider and largest machine in the world, serves as a means to test certain theories of particle physics. Several goals of the project relate to understanding subatomic particles that propel jets. A jet may arise from particles called quarks, gluons, W-bosons, top-quarks, or Higgs bosons. Understanding jets at the subatomic level helps particle physicists understand fundamental forces in the universe.

Here’s where data science comes in: to understand subatomic jet particles, physicists need to be able to classify these particles. Different types of neural networks can address this classification problem. In recent research, recursive neural networks have been used based on an analogy between quantum chromodynamics (QCD) and natural languages — as a sentence is composed of words following a syntactic structure, a jet is composed of particles following a structure determined by quantum chromodynamics.

But in new research from Joan Bruna, Kyunghyun Cho, both CDS Faculty and Assistant Professors of Computer Science and Data Science, Kyle Cranmer, CDS Affiliated Faculty and Associate Professor of Physics, Isaac Henrion, and Johann Brehmer, jets are represented as graphs and studied with a Message Passing Neural Network (MPNN) rather than a recursive neural network. The researchers applied their MPNN to a binary classification problem, attempting to distinguish between two classes of jets: QCD jets (arising from a known mixture of quarks and gluons) and W jets (arising from bosons decaying into two quarks).

Compared with the performance of recursive neural networks in achieving this classification, MPNNs did significantly better. Recursive neural networks are restricted to a tree structure while graph-representation for MPNNs allows information between all particles to be exchanged. However, the researchers caution that model configuration must be carefully designed for MPNNs because it influences the final result.

MPNNs offer a new data-driven avenue to explore jets compared to traditional jet clustering algorithms. For future research and experimental design, the researchers suggest improvements for MPNNs to avoid expenses due to high complexities and advances to handle larger scale inputs.

By Paul Oliver