Forschungsprojekt ::

Uncertainty-aware and physics-informed machine learning for short range atmospheric forecasts

Projektbeschreibung

Quantification of evolving uncertainties is required for both probabilistic forecasting and data assimilation (DA) in numerical weather prediction (NWP). In current practice, an ensemble of NWP model simulations is often used as a primary tool to describe these uncertainties. In this proposal, we explore an alternative approach that is motivated by recent developments in machine learning (ML) research, namely, uncertainty quantification and physics-informed ML. As to the former, new ML methods are currently being developed that are able to represent (predictive) uncertainty in an appropriate manner, with a specific emphasis on distinguishing deferent sources and types of uncertainty, such as aleatoric and epistemic. As to the latter, physics-informed ML seeks to increase the data efficiency and transparency of purely data-driven black-box modeling through the incorporation of physical constraints. The goal of this project is to develop tailored algorithms that will ensure the physical plausibility and proper quantification of the predictions’ uncertainty forward in time. This comes with various challenges, notably the adaptation and extension of existing uncertainty-aware ML for the specific type of forecast atmospheric variables, the design of physics-informed ML methods tailored to this application, as well as the combination of these two.
In a first part of the project, we plan to use an idealized two-dimensional model. We will start by following the propagation of initial perturbations in a parameter through time for all model variables. The explicit use of physical laws in ML will make the algorithms interpretable. In addition, adequate DA methods corresponding to the new ML models will be developed, with the additional benefit of improved online learning. The second part of the project will build on the ensemble generated by a high resolution numerical model. Here, we also first focus on the evolution of uncertainty due to uncertainty in a parameter (e.g., perturbed condensation cloud nuclei) using methods developed in the first part of the project. Results of ML will be compared when the source of uncertainty is changed, for example, due to use of a different microphysics scheme. The ability of ML to distinguish different sources and types of uncertainty (aleatoric and epistemic) will be tested. Moreover, the new ML methods will be compared to the ensemble approach for both models.