Machine learning, and deep learning in particular, is becoming increasingly central to geophysical data assimilation, where it can enhance, complement, or even replace parts of the classical data assimilation cycle. I will illustrate recent advances in the field through two examples.

In the first, we investigate how deep learning can be used to discover new optimal analysis operators for sequential data assimilation applied to chaotic dynamics. I will show that the resulting “data assimilation networks” can address some of the most important challenges in classical data assimilation, such as properly representing the analysis uncertainty and accounting for departures of background statistics from Gaussianity.

In the second example, I will show how a generative AI-based surrogate model can be naturally incorporated into a classical yet advanced ensemble variational data assimilation scheme. The aim is to account for a broader class of model errors than those represented, for instance, in weak constraint 4D-Var.
Both cases will be illustrated through data assimilation experiments with low-order dynamics, potentially involving non-trivial model errors.