GRAIL: A Hardware-Aware Machine-Learning Pipeline for On-Board Classification of Space-Calorimeter Events
Computing Conference (CC 2026) – published in Intelligent Computing, Lecture Notes in Networks and Systems, Springer.
Abstract
A study of GRAIL, a hardware-aware machine learning pipeline for dimensionality reduction and onboard classification of calorimetric events in space astrophysics missions, designed for embedded and edge-AI platforms. The work compares several dimensionality reduction techniques (PCA, ICA, UMAP, variational autoencoders, XGBoost-based feature selection) and classifiers, and identifies XGBoost combined with CatBoost as the best trade-off between accuracy and computational cost. Deployment tests on edge platforms (Arduino Uno, NVIDIA Jetson Orin Nano) using Geant4-simulated data show that onboard inference is feasible, while also highlighting the current limitations for space qualification.
Authors
A. Garinei, A. Vispa, S. Speziali, M. Proietti, F. Fallucchi, S. Meola, A. Santiangeli, F. Zuccari, E. W. De Luca, U. Di Matteo, R. Giuliano, M. Marconi, E. Piccioni, M. Bossa
Link
Publication date: 01/06/2026 Last updated: 02/10/2026 17:21