The Work
Reichelt built CloudFlow, a conditional flow-matching model that generates ten-channel synthetic satellite data at 1 km resolution measuring key cloud properties from coarse 25 km atmospheric inputs. To probe temporal generalisation of the model, there is a 7 year gap between the training data and the held-out test set.
What Worked
CloudFlow matched observed power spectra to high wavenumbers and delivered the best MAE and CRPS on most channels, roughly 20% better CRPS on the radiance bands.
What Has Been Achieved
Its attention-encoder variant matched observed power spectra to fine scales, produced a cloud-top-height error on the same order as the satellite retrieval itself, and delivered the best probabilistic scores on most channels.
What Was Learned
Learned encoders were expected to dominate hand-designed cloud controlling factors, but the classical factors stayed a surprisingly strong baseline, a sign they already capture much of the first-order physics.
About
Tim holds a PhD in ML and Statistics from Oxford. He was previously a Postdoc in the Climate Processes group at the University of Oxford, developing novel compression algorithms for atmospheric data. He has broad interests in probabilistic machine learning, Bayesian statistics, and deep learning.
Lab & Advisors
Philip Stier
Oxford University
Professor of Atmospheric Physics
Philip Stier leads the Climate Processes Research group at Oxford University, specializing in aerosol-cloud interactions, cloud feedbacks, and climate modeling. His work is pivotal in addressing uncertainties related to clouds in climate prediction and has significantly contributed to global atmospheric research initiatives.
