Gunnar Behrens

Timeline

1. 2022

   Variational encoder for convection

   In 2022, he is first author of a paper with Tom Beucler, Pierre Gentine, Fernando Iglesias-Suarez, Michael Pritchard and Veronika Eyring. It uses a Variational Encoder Decoder, a non-linear dimensionality reduction technique, to learn convective processes in an aquaplanet superparameterized climate model simulation, and the model reproduces them with only five latent nodes.

2. 2023

   Co-authoring the ClimSim dataset

   In 2023, he is a co-author of ClimSim, a NeurIPS Datasets and Benchmarks Track paper that presents 5.7 billion pairs of multivariate input and output vectors for hybrid ML-physics research. A consortium of climate scientists and ML researchers develops the multi-scale climate simulations behind it, and the dataset is global in coverage and spans multiple years.

3. 2023

   A NeurIPS outstanding dataset award

   In December 2023, the NeurIPS organizers announce ClimSim as an Outstanding Paper in the dataset category of the Datasets and Benchmarks Track. He is among the authors listed on the winning paper.

4. 2024

   A Bremen and DLR dissertation

   In 2024, he completes the dissertation "Understanding and Modelling Convection with Machine Learning" at the Universität Bremen and the Institut für Physik der Atmosphäre at the German Aerospace Center (DLR). It develops a variational encoder decoder that separates convective regimes and an ensemble approach that represents convective processes better, and it shows that the ensembles' uncertainty quantification stabilizes hybrid simulations and reduces biases.

5. 2025

   Multi-member parameterizations in JAMES

   In 2025, the Journal of Advances in Modeling Earth Systems publishes his paper on deep learning multi-member and stochastic parameterizations, written with seven co-authors including Tom Beucler, Pierre Gentine and Veronika Eyring. The parameterizations learn subgrid convective and turbulent processes of a superparameterization embedded in an Earth System Model, and the multi-member version improves the representation of convective processes in the planetary boundary layer compared to individual neural networks.

6. July 2026

   Preprint on diurnal convective cycles

   In July 2026, a preprint with Lara Leko, Patrick Eriksson and colleagues is posted for discussion on EGUsphere with Behrens as second author. It uses the Chalmers Cloud Ice Climatology, a machine-learning satellite product, to compare the diurnal cycles of frozen water path and precipitation in satellite observations, ERA5 and DYAMOND km-scale models.

7. September 2026

   Co-authoring the ICONEval framework

   In September 2026, Geoscientific Model Development publishes ICONEval v1.0, an open-source evaluation framework built on ESMValTool for the ICON model, with Axel Lauer as first author. He is one of its 14 authors, and the framework groups sanity checks, physical consistency checks and climate fidelity diagnostics, including for hybrid ML-enhanced Earth system models.

8. As of October 2026

   Postdoc at Chalmers

   By 2026, he moves from the University of Bremen's Institute of Environmental Physics to the Department of Environmental and Energy Sciences at Chalmers University of Technology in Gothenburg, Sweden. There he is a postdoc in Geoscience and Remote Sensing.