Jet charge identification with Transformers in the ATLAS experiment

Identifying particle jets by flavor is an important component of many analyses in the ATLAS experiment, especially those involving final states with bottom and charm quarks. Jet flavor identification in the experiment has recently been revolutionized using Transformer-based neural networks with an approach characterized by the inclusion of physically motivated auxiliary training objectives. We summarize these recently developed techniques and present an extension of our jet flavor identification algorithm that infers the electric charge of the initiating quark. This is done via an additional auxiliary task, which allows flavor and charge classification to be performed simultaneously and provides a powerful discriminant for processes with charge correlated final states. Once deployed, this improved algorithm could provide significant sensitivity improvements across multiple analyses, such as measurements of H(cc).