The LAFI consortium will generate large amounts of observations and model outputs of the L-A system with unprecedented detail and resolutions. However, understanding the processes and feedbacks underlying this complex data is a very challenging task. The atmospheric flow is chaotic in nature. The actual interactive complexity of turbulence dynamics within the atmospheric boundary layer and entrainment fluxes is still poorly understood. The highly non-linear interaction between many variables of the L-A system is therefore very difficult to describe.
In recent years, however, with deep learning (DL) techniques, methods were developed which are specifically designed to identify hidden internal structures from large and complex datasets. In the framework of LAFI, these capabilities of DL will be exploited to extract the systematic interrelations underlying the processes and feedbacks in the L-A system. More precisely, we will 1) develop hierarchies of machine learning approaches and apply importance weighting to identify key driving parameters and variables; 2) enhance DL and other machine learning techniques with physics-informed mechanisms; 3) include L-A system-informed inductive biases in DL, and 4) develop DL-based foundation models, which will target the inference of general, L-A state characterizing structures. In this connection, we aim not only for improved representations of relationships or identification of structures but also for an improvement of process understanding.
Fig. 1: Schematic of a foundation model:
