In this article , Karl Otness and co-authors present a new multiscale machine-learning approach designed to improve predictions in dynamical systems. The method captures information moving both from fine to coarse scales and from coarse to fine, boosting model accuracy and stability, with only minimal added computational cost compared to standard architectures. The team evaluates the approach on an idealized fluid-dynamics closure task, where the multiscale networks learn to correct a chaotic model by representing unresolved small-scale processes. The work highlights the potential of multiscale AI architectures to enhance the reliability of physical system modeling.