Machine learning approaches and new data science algorithms are an emerging frontier for the atmospheric sciences. We explore whether newly developed physics-guided machine learning algorithms trained with atmospheric model data or observations can serve as emulators for physical processes in weather and climate models, such as the time-consuming solar radiation code, precipitation mechanisms, or the shallow or deep convection cloud schemes. A second, less aggressive approach is to utilize machine learning approaches for the estimation of uncertain parameters in the subgrid-scale physical parameterizations of atmospheric models. We use idealized weather and climate model configurations to intercompare the pros and cons of various machine learning algorithms, such as linear regression, random forests, boosted forests, artificial neural networks and deep neural networks with and without convolutions. In addition, we are interested in machine learning approaches to understand and foster the predictability of the climate system over subseasonal-to seasonal (weeks-to-months) time scales.