Speaker
Description
The integration of advanced artificial intelligence techniques into astroparticle and high-energy physics experiments represents a transformative shift in both data analysis and experimental design. As space missions become increasingly sophisticated, the adoption of AI-driven methodologies is essential for optimizing detector performance and ensuring robust scientific results.
In this work, we present the development and application of machine learning approaches for two key tasks in space-based experiments: particle tracking and calorimetric shower classification. For tracking, we introduce innovative algorithms based on Graph Neural Networks (GNNs), a central paradigm of geometric deep learning that naturally exploits the graph structure of detector data, where nodes represent energy deposits (hits) and edges encode spatial and temporal correlations. A major challenge in this context is the highly noisy environment, dominated by backscattering hits from the calorimeter, which complicates the reconstruction of primary particle trajectories. To address this, we propose a dedicated GNN-based node-classification pipeline designed to effectively discriminate noise hits from true signal hits, significantly improving tracking accuracy.
In parallel, we investigate AI techniques for the classification of electromagnetic and hadronic showers in space calorimeters (transformer and boosting algorithms). Using Monte Carlo simulations and dedicated feature engineering, our results demonstrate the strong potential of AI methods to enhance classification performance in complex detector environments.
To manage the large data volumes and computational demands of these tasks, we implement a distributed training strategy with parallelization across multiple GPUs. Extensive tests conducted on NVIDIA A100 and H100 architectures at the ReCaS data center, the Leonardo infrastructure and the Napoli GPUs cluster show substantial reductions in training time and memory usage, while preserving model stability and performance.
Overall, this study highlights the effectiveness of modern AI techniques, particularly GNNs and scalable training strategies, in addressing key challenges in space-based particle physics, paving the way for more efficient and accurate data analysis in future experiments.