GraphSAGE Classifier for Top Quark Tagging

GraphSAGE-based Graph Neural Network for identifying physics events originating from top quark decays.

n high energy particle physics experiments, such as those at the Large Hadron Collider, the amount of physics events is typically far too large to feasibly record it all, and the majority of events are from well-understood processes that aren't useful for discovering interesting physics such as Higgs Boson or Top Quark decays. Therefore, it is necessary to have a trigger system to quickly analyze incoming data and determine whether it is worth keeping. This is performed by hardware-level triggers (typically some kind of electronic circuit) and software-level triggers (algorithms which analyze data).

Top quark decays are one such example of an "interesting" physics process that we may wish our trigger system to select for. Top quarks are the heaviest of the six flavors of quarks, and their decays can be difficult to reconstruct. Its short lifetime means that it decays before it can hadronize, meaning it always decays into lighter quarks first which then form hadronic jets. Thus, its signature could be confused with that of the lighter quarks. Machine Learning can be used to train an algorithm to distinguish between the subtle differences. We call this task "top-tagging."

In this demo, some Monte Carlo (simulated) data containing Top Quark decays (signal) and lighter quark or gluon decays (background) is used to train two different neural network classifier models for top-tagging. The first model is a ResNet-based Convolutional Neural Network, trained on image representations of the events. The second model is a GraphSAGE-based Graph Neural Network, where the data has been converted to a graph representation.