Luís Simões
Marília Simões
Nikki Grens
ML-Analytics

Abstract
Machine-learning pipelines are currently being developed to generate highly processed data products for the Ariel mission, offering substantial computational advantages over traditional pipelines. However, such deployments in science-mission contexts remains constrained by concerns regarding reliability, robustness, and trustworthiness. This talk presents a study that investigated the design and evaluation of a safety cage architecture to supervise machine-learning predictions in the context of the ESA Ariel mission. The "safety cage" architecture is a supervisory layer designed to monitor the operational range of spectroscopy models. By integrating multiple indicators, including conformal prediction, out-of-distribution detection, and ensemble consistency, the framework identifies when a model is extrapolating beyond its validated domain. We demonstrate the methodology using simulated Ariel datasets, showing how this approach enables controlled degradation rather than catastrophic failure. This work provides a systematic path toward building trustworthy, autonomous inference pipelines for the next generation of space science missions.

2026 May 14, 13:30

IA/U.Porto
Centro de Astrofísica da Universidade do Porto (Auditorium)
Rua das Estrelas, 4150-762 Porto