Estimating Wind Fields Using Physically Inspired Neural Networks With Aircraft Surveillance Data

Estimating wind and reconstructing wind fields are important for aviation, as accurate wind information can enhance aircraft flight planning and safety. In this study, a physically inspired neural network is designed to rapidly estimate wind fields from scattered local wind measurements obtained from aircraft. To train the network, we generate a large synthetic set of data from ERA5 reanalysis data based on the actual flight paths. Next, to apply the model, actual aircraft measurements are used as input, and the wind field over the entire airspace at each altitude is reconstructed. The network adopts a new physical loss function, which can smooth the predicted flows. The network can predict flow fields on both simulated and real measurements, given sufficient input data. This approach improves upon traditional methods, such as the numerical forecast model and our previous Meteo-Particle model, resulting in a significant reduction of wind magnitude error by 40% to 2.85 m/s and directional error by 27% to 11.2 degrees. Although the error metrics fluctuate depending on the nature of the flow, the network is suitable for nowcasting and short-term forecasting based on limited wind observations.