Atmospheric Science commonly deals with complex spatiotemporal fields. Methods based on deep learning (DL), such as convolutional neural networks, have turned out to be powerful tools for analyzing such data. However, predictive DL models are typically trained to optimize loss functions such as the root-mean-square error, which leads to blurry predictions and does not give a quantitative estimate of the uncertainty of the prediction, whose importance is particularly emphasized in the weather and climate fields. Alternatively, probabilistic losses such as cross entropy can produce pointwise uncertainties, but still fail to represent the spatial correlations in the uncertainty.
Generative models are able to produce diverse, realistic samples. This makes them – and especially their conditional variants – well suited for representing uncertainty through sample diversity. In the recent years, generative adversarial networks (GANs), have found applications in weather and climate data processing. They can be used for common problems in this field, such as generating physical fields from the corresponding in-situ and remote sensing observations, increasing the resolution of observed data, or predicting the time evolution of data fields.
In this presentation, I will give an overview on the applications of generative models in the atmospheric science, with an emphasis on my own work in processing cloud and precipitation observations with them. I will also discuss more generally which problems in climate science could (or already do) benefit from generative models. Furthermore, I will discuss the current challenges and open questions for training generative models for weather and climate applications, and in validating and interpreting their results.