Physics-informed AI systems - how the constraints from NWP can support development of better AI models

AI-based methods have already proven their skill in weather prediction. The availability of high quality training data and ever more advanced model architectures opened the door for a new range of models providing predictions fast and at a low operational cost. Further development of such models, apart from seeking advancements in the architectural design and increased quality of the training datasets, might require a new approach. One still under-explored dimension is to tighten the link between the data-driven (often weather-system-agnostic) methods and the well established knowledge of atmospheric physics represented by NWP. 

Physics-awareness in AI model development may guide training, reduce compute requirements and improve the consistency of the predicted variables. In this talk we discuss the possible approaches to physics-informed AI model design, ranging from physics-based terms in the cost function to hybrid physical-neural architectures. We show the impact these methods have on the training process and discuss possible improvements in the forecast skill and physical consistency.