When using AI for control and optimization of critical infrastructure, there is a risk that the underlying (learned) models do not cover all aspects of the given plant, and henceforth the modelled system may enter a potential unstable state.

In this talk, we present the challenges in more detail, and we introduce the notions of shielding and learning under shielding. We couple the notions to the exploitation of digital twins for identifying (stochastic) parameters and for setting up virtual test environments.