Physics-informed Machine Learning for Data Assimilation in High-Dimensional Space Weather Models

Space Weather (SW) is represented by multiple high-dimensional dynamical systems coupled together and driven by solar activity. However, predictions from state-of-the-art physical models often differ from the observational data. The differences between model output and observations are attributed to 1) incomplete knowledge of the physical and dynamical processes and 2) uncertain inputs such as key model parameters, drivers, and initial and boundary conditions. Prediction capabilities can be significantly improved by guiding the models with observation data. However, existing methods for model-data fusion (e.g. data assimilation) make assumptions that sometimes prove inefficient and ineffective at extracting information from observation data when used with complex (large-scale and high-dimensional) nonlinear models, resulting in prediction inaccuracies. Thus, there is a widely recognized and critical need across the space weather for new perspectives on combining complex and nonlinear models with observation data for (i) a better understanding of the dynamical processes, and (ii) facilitating better predictions of those processes.

This talk will discuss physics-informed machine learning as a means for reducing the complexity or dimensionality of high-dimensional system models. The reduced order models encapsulate state-of-the-art physical knowledge learnt through simulation data. Traditional approaches for data assimilation can then be used for integrating observations with the model towards improving forecasts. We will demonstrate the methodology with two different applications: thermosphere and ground magnetic perturbations (dB/dt). Finally, we will present the vision for a new end-to-end, data-driven framework for probabilistic space weather modeling and forecasting (‘Diet-SWMF’) where the components are ideally physics-informed, machine learned models but can also be purely analytical/statistical.