Earth system and climate models are fundamental to understanding and projecting climate change. Although they have improved significantly over the last decades, considerable biases compared to observations and uncertainties in their projections still remain. We will take a new approach by harvesting output from high-resolution cloud resolving models to develop new machine learning-based parametrisations for small-scale processes (e.g. clouds) that cannot be explicitly resolved in global climate models. The new ML-based climate model will have the potential to eliminate some of the long-standing systematic errors and to provide more robust climate projections. The project will develop: 1) new architectures for learning feature representations of the subgrid processes, 2) estimates of stability of the networks before the more ambitious far-end goal of coupling deep learning with climate models, and 3) new losses that ensure physical consistency of the networks and avoid violations of the most fundamental laws of physics. This work is done in close collaboration with Prof. Pierre Gentine (Columbia University, New York) and Prof. Markus Reichstein (Max Planck Institute for Biogeochemistry, Jena) as part of the European Research Council (ERC) Synergy Grant “Understanding and Modelling the Earth System with Machine Learning (USMILE)”.