Speaker
Description
Calorimeter calibration converts raw detector signals into estimates of incident particle energy. In highly segmented systems, the detector response depends not only on deposited energy but also on light collection efficiency, sensor response, shower topology, and local geometry. These effects are particularly significant in high-granularity calorimeters with many independent readout channels. The HG-DREAM calorimeter provides complementary Cherenkov and scintillation signals, offering detailed information about shower development. We introduce a modulation-based convolutional neural network (mod-CNN) that extends traditional linear calibration methods. While linear regression reconstructs energy as a fixed weighted sum of channel signals, it cannot capture nonlinear effects such as saturation and position dependence.
The mod-CNN preserves the weighted-sum structure but introduces a bounded, event-dependent modulation derived from the spatial and amplitude patterns of each event. This enables adaptive channel weighting that reflects the detector response. The method improves energy reconstruction performance and provides a diagnostic framework for identifying nonlinear sensor behavior and response variations. Current results show improved electromagnetic energy response and resolution compared to a linear baseline.