Machine Learning (ML)-based parameterizations of atmospheric convection have long been hailed as a promising alternative, with the potential to yield higher accuracy at a fraction of the cost of more explicit simulations. This work , co-authored by Mohamed Aziz Bhouri, investigates the coupled out-of-distribution extrapolation capabilities in “online” testing of different ML-based parameterization designs of atmospheric convection. Their results show that these design decisions are not enough to obtain satisfactory generalization benefits. The obtained improvement indicates the necessity of using multi-climate simulated training data.