Simulations and analysis for next-generation CMS calorimetry with nanophotonic and neuromorphic readout technologies

Future particle-physics experiments, together with upgrades to existing facilities such as CMS at the LHC, will require detector systems capable of preserving detailed spatial and temporal information under strict constraints on bandwidth, latency, power consumption, and hardware complexity. Within this context, the PHINDER project investigates an integrated architecture combining scintillation-photon detection, optical communication, and neuromorphic processing. This thesis addresses two related objectives within PHINDER. The first is the development of an interpretable Monte Carlo model of scintillation-photon production and transport, intended as a physical reference for a future differentiable implementation. The second is to assess whether the arrival-time distribution of detected photons contains spatial information complementary to the integrated photon yield and conventional detector segmentation. A sampling calorimeter composed of tungsten absorbers and plastic scintillator was used as a controlled test bed. Energy deposits generated with Geant4 were converted into optical photons and propagated with a dedicated stochastic simulator modeling photon production, transport, and collection in time, position, and direction. The resulting signals were analyzed in local 2×2 groups of neighboring cells. For each cell, reconstruction used both the total detected-photon count and the photon arrival-time distribution. A hybrid multi-task architecture combining spiking and conventional deep-learning components was developed. LIF-based branches encoded temporal signals using cell-relative and group-relative time references, while a dense branch processed photon counts. The representations were fused through multi-head self-attention to reconstruct cell activity, deposited energy, the transverse energy-weighted centroid, and track multiplicity.