Speaker
Description
We present a convolutional neural network (CNN) for real-time jet substructure identification in the ATLAS hardware trigger at the High-Luminosity LHC (HL-LHC). This approach introduces substructure discrimination at Level-1, enabling efficient separation of jets from hadronically decaying boosted objects (e.g. W/Z/H bosons and top quarks) from the dominant QCD background under extreme pileup conditions. The CNN operates on jet “images” built from NxN grids of trigger towers or calorimeter cells centered on jet seeds.
We evaluate performance using inputs from both the existing Global Feature Extractor (gFEX) and the future Global Trigger systems, and demonstrate robustness across multiple jet definitions, pileup mitigation strategies, and calibration schemes currently under development for HL-LHC. Hardware feasibility is assessed through detailed studies of latency and resource usage on FPGAs and emerging AI-accelerator architectures (e.g. AMD Xilinx AI Engines).
This study demonstrates the deployability of deep-learning-based jet substructure tagging at the hardware trigger level. The resulting rate reduction for selections requiring both jet pT and substructure enables significantly lower thresholds for large-radius jets, substantially improving sensitivity to boosted hadronic signatures.