Subgrid-scale vertical velocity variability (σw) plays a key role in aerosol activation, cloud droplet number concentration (CDNC), and cloud microphysical evolution. Despite its importance, σw remains one of the highly uncertain parameters in climate models, particularly under extreme atmospheric conditions. Recently developed machine-learning turbulence parameterizations, trained on global high-resolution climate model simulations, offer a promising alternative to traditional schemes. However, their ability to apply and generalize to real atmospheric conditions and to regimes that differ substantially from those represented in the training simulations and the resulting implications for cloud processes remain largely untested.Here, we evaluate the deep-learning based σw parameterization, Wnet across two physically contrasting observational regimes that are highly relevant for aerosol–cloud interactions: (i) the ultra-stable Arctic boundary layer observed during the CLAVIER campaign at the Villum Research Station, and (ii) strong orographic turbulence associated with cloud formation during the CHOPIN campaign at Mt. Helmos in the Mediterranean. Using high-frequency observations, we drive Wnet in offline mode and compare its σw estimates against observed vertical-velocity variability, with a focus on conditions controlling cloud activation. We further assess the robustness of Wnet by diagnosing out-of-distribution (OOD) atmospheric states relative to its global training space and examining how such states are associated with systematic σw errors. This framework enables identification of distinct ML failure modes under different atmospheric conditions, thereby elucidating the physical boundaries of applicability of ML-based Wnet turbulence scheme. Finally, we investigate when and where σw errors translate into meaningful biases in cloud-relevant quantities, particularly CDNC, by linking σw discrepancies to observed cloud properties and activation regimes. By explicitly connecting ML-driven turbulence errors to cloud microphysical impacts, this study provides a physically grounded evaluation of ML turbulence parameterizations in regimes critical for aerosol–cloud interactions. The results will inform the safe and interpretable use of ML-based σw schemes in Earth-system models and highlight key challenges for their application in extreme atmospheric environments.