Falga et al. present a new machine-learning–based parameterization for turbulent momentum fluxes that works consistently across both oceanic and atmospheric boundary layers. Trained on large-eddy simulations, the neural network captures key turbulent features missed by traditional schemes and significantly improves boundary-layer wind predictions in climate models, reducing errors by a factor of 2–3 under convective conditions. The approach is robust to surface flux biases and generalizes well beyond the training data, highlighting the promise of unified, data-driven turbulence closures for next-generation climate models.