Granular Calorimetry Integrating Neuromorphic Computing Readout for Future Colliders

The increasing data rates and power constraints of present and future high-energy collider experiments demand readout and inference systems that go beyond the capabilities of conventional von Neumann computing architectures. This thesis investigates the application of neuromorphic computing to the reconstruction of physical quantities from calorimeter signals, with the aim of demonstrating the viability of Spiking Neural Network as a low-power, low-latency alternative to conventional signal processing at the detector front end. The system under study consists of a homogeneous lead tungstate (PbWO4) calorimeter block, simulated using the GEANT4 framework, in which 100 GeV protons, charged pions, and charged kaons initiate hadronic showers. The resulting scintillation light is propagated to an array of idealised photosensors instrumented on each calorimeter cubelet, and the recorded photon time series are encoded into binary spike trains via a threshold-based multi-channel scheme. A fully connected SNN, comprising two hidden layers of LIF neurons, is trained via surrogate gradient backpropagation to regress three classes of physical observables directly from the spike trains: the total deposited energy, the energy-weighted spatial centroid of the shower, and the energy-weighted spatial variance of the deposition pattern. The results demonstrate that the neuromorphic system is capable of inferring topological shower information ordinarily associated with highly segmented calorimeters, yet obtained here from a single homogeneous detector block without fine physical segmentation of the active medium. This establishes a proof of concept for neuromorphic calorimetric reconstruction and identifies a set of promising directions for future work, including learnable spike encoding, convolutional and recurrent architectures, full-calorimeter reconstruction, and particle identification. A potential hardware realisation based on III–V semiconductor nanowire photonic components is also discussed, which could enable ultra-fast, sub-nanosecond inference integrated directly into the detector readout chain.