For Machine Learning models to be valuable tools in the study of future exoplanet observations, they need to reliably operate in contexts where covariate and concept drifts are guaranteed, and may even in some instances be desired!

Efforts on ML Operations for the Ariel telescope focus on the concerns of how to reliably serve predictions to the scientific community under such challenging conditions.

This talk will provide an overview of directions being pursued at mlanalytics.ai to bring models to a production grade status in the mission, such as data curation and evaluation, domain adaptation, out-of-domain detection, uncertainty quantification, and safety cages for operational range bounding.