Physic-Informed AI for Wind Forecasting from Satellite Observations

AI-based weather models have recently shown good skills in predicting the state of the atmosphere
simply by learning from the data. The complexity of the atmosphere can be captured by models with
a very large number of parameters and trained on large volumes of data. This has prompted different
meteorological offices (e.g., the ECMWF) to develop their own weather AI models. The results show
that they can be very effective to prevent extreme events, however they are sometimes deficient when
it comes to events involving complex physics.

We aim to overcome this drawback by training these AI models with physical laws. This is the
principle behind the development of PINNs (Physic-informed neural networks). This would allow to
use the power of AI models to extract information from data and guide the reuse of this information
so that it respects physical laws.

The final objective of our study is to use these architectures to predict low-level wind shear. This
phenomenon is a hazard that can seriously disrupt aircraft in approach/departure phases and even
cause accidents. We will begin a reconstruction of the wind field using observations from weather
stations, then build a nowcast (0-2h) / forecast from this initial state. Weather station observations are scarce
(low spatial and temporal resolution) and subject to errors (instrumental, due to the environment, etc.).
High-resolution wind nowcasting/forecasting based solely on these observations is therefore really challenging. We
will use PINN architectures to regularize these observations and extract the information needed to
produce a high-resolution short-term forecast.

Where many of the leading data-based weather models are trained on ERA-5 and take ERA-5/operational forecasts as their initial conditions. This means that the data assimilation step is always present (expensive) and that the forecast is based on a forecast from a previous run. By working directly from the observations, we can avoid the classic data assimilation step and avoid introducing a bias (linked to the analysis) on our initial condition.