Secondary Vertex Reconstruction at The High-Luminosity LHC with MaskFormer Architecture

With the upcoming era of the High-Luminosity LHC, secondary vertex reconstruction will face significant computational challenges in dense environment, highlighting the need for novel and efficient identification algorithms. This report introduces
a new approach, adapting the transformer-based MaskFormer architecture from computer vision to reframe SV finding as an end-to-end instance segmentation problem. A dedicated data processing pipeline was developed to convert simulated
ROOT data into a uniform HDF5 format suitable for a multi-task model designed to simultaneously perform SV classification, track assignment (masking), and vertex property regression. The model, trained on a realistic, class-imbalanced dataset,
demonstrated excellent performance on the classification and track assignment tasks, achieving results comparable to a baseline trained on a perfectly balanced reference sample. However, the vertex regression task proved highly sensitive to this imbalance, showing significant underperformance compared to the reference. These findings establish the query-based transformer architecture as a powerful and viable new direction for tackling the combinatorial challenge of vertex reconstruction. The results also highlight that achieving precise regression in a multi-task setting on imbalanced data is a key challenge, suggesting that future work must prioritize the optimization of the loss function weights and target scaling strategies to unlock the full potential of this approach.