Machine Learning Methods for Wind Profile Recovery in the Atmospheric Boundary Layer

ICLR 2026 Conference Submission19654 Authors

Keywords: Wind Prediction, Climate, Tabular Data, Neural Operator

Abstract: Reconstructing atmospheric boundary layer wind profiles is crucial for weather prediction and wind energy applications. We study the task of predicting vertical profiles of the zonal (east–west) and meridional (north–south) wind components from the Integrated Global Radiosonde Archive (IGRA), given geostrophic wind and auxiliary predictors such as season, time of day, temperature difference, and pressure. We propose several machine learning architectures for this task, including CatBoost, TabM, and FT-Transformer, against classical baselines such as the power-law profile and Monin–Obukhov similarity theory. On the IGRA dataset, modern ML models achieve reconstruction errors of about 1.5 m/s for both wind components, which is superior than analytical models.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 19654