AI-Accelerated Signal Reconstruction for the ATLAS Tile Calorimeter at the HL-LHC

The ATLAS Tile Calorimeter (TileCal) currently reconstructs energy using a linear signal-processing method. For the Phase-II HL-LHC upgrade, readout will move to a 9-sample sliding window, where this approach becomes less reliable because many proton-proton collisions overlap in time. This project develops compact neural networks (MLPs and 1D-CNNs) that improve energy reconstruction while remaining small enough for real-time FPGA deployment. The approach combines a floating-point teacher CNN trained offline, knowledge distillation into a compact student model, quantization-aware training with 1-8 bit mixed precision, and magnitude pruning, targeting a strict 225 ns algorithmic latency within the 1.7 μs TilePPr budget. Deliverables include reproduced linear baselines on Phase-II simulation data, a trained teacher CNN with benchmarked energy resolution, a distilled and quantized student model, a prototype offline reconstruction integration, and a full FPGA synthesis report with latency and resource usage on a Xilinx Ultrascale+ target device.