A large number of machine learning studies are currently performed in Earth System science to extract information from observations or model data, to learn the dynamics of components of the Earth system, or to build tools to improve weather and climate predictions (for example in model post-processing). These studies require both the domain knowledge of Earth system scientists and the expertise of machine learning experts. The astounding success of machine learning in other scientific disciplines – such as computer vision, speech recognition, robotics, and autonomous driving – is not least a result of the free and open sharing of benchmark datasets and software, which allows easy reproducibility by the community as well as the quantitative comparison of the quality of machine learning solutions. Contrary to many other machine learning disciplines, Earth system datasets and validation methods require thorough documentation to achieve the goal of providing a fair and firm basis for model comparisons according to the accepted community standards. Therefore, the Earth system science community has expressed a need for dedicated benchmark datasets and machine learning tools to process Earth system data.

With this special issue, we offer a platform to develop and openly share these benchmark datasets together with specific machine learning tasks and evaluation metrics as well as innovative solutions for machine learning applications in Earth system science with the wider community of machine learners and environmental scientists.

The special issue is managed jointly by ESSD and GMD and invites manuscripts focusing either on the description of new benchmark datasets (ESSD) or new machine learning model developments (GMD). Dataset papers can reference model papers and vice versa, but this is not a requirement as long as both the datasets and model codes are made freely available and easily accessible. To be accepted as special issue contribution a manuscript must define a specific machine learning task including a loss function and evaluation metric and the authors must provide executable code which demonstrates the machine learning problem with a vanilla solution. The usual guidelines of both journals apply.