Speaker
Description
Accurate modelling of particle detection in hybrid semiconductor detectors remains a complex task, particularly when aiming to reproduce detailed sensor and readout responses. While modern simulation frameworks such as Allpix Squared provide a solid foundation, subtle differences between simulated and experimental data can persist, especially in fine structural features of particle tracks.
In this work, we extend the Allpix Squared framework with a machine learning-based model to enhance the realism of simulated data for proton beams at energies of 70, 150, and 225 MeV. The machine learning model is trained on data acquired at multiple incidence angles (0°, 30°, 60°, 75°), while the 45° configuration is intentionally excluded from training and used to evaluate the model’s ability to generalize to unseen geometries.
Our approach learns to refine spatial and intensity-related characteristics of simulated detector responses without requiring explicitly paired datasets. The results demonstrate that the machine learning extension of the Allpix Squared framework captures underlying physical features, such as deposited energy, cluster size, and cluster height, and generalizes well to unseen angular configurations.
This approach provides a flexible framework for generating realistic synthetic datasets, supporting detector characterization and contributing to the development and validation of advanced simulation tools.