AI for Climate Science Seminar Series: Climate Emulator applications in the context of Destination Earth

For our October edition we are excited to announce a online lecture by Alberto Sanchez-Marroquin and Marvin Axness Ferrando from the Computational Earth Sciences Group of the Barcelona Supercomputing Center.

In this seminar, Alberto Sanchez-Marroquin and Marvin Axness Ferrando will present insights into Climate Emulator applications in the context of Destination Earth, showcasing a set of ML-based applications developed at the BSC, built on top of datasets from the Destination Earth Climate Digital Twin.

Alberto Sanchez-Marroquin is a multidisciplinary researcher and data scientist at BSC, with years of experience in Atmospheric Science and Machine Learning. He holds a PhD in Atmospheric Science from the University of Leeds and a degree in Physics. 

Marvin Axness Ferrando is an HPC Research Engineer at BSC, working with Destination Earth datasets and Climate Emulators. He holds an MSc in Nuclear Physics and a degree in Physics. 

For online participation, a registration is necessary.
Please register here.

First presentation: "A Climate Emulator for Destination Earth" by Marvin Axness Ferrando

Destination Earth (DestinE) is an initiative of the European Commission to build a highly accurate Climate Digital Twin of the Earth system, enabling the exploration of past, present and future climate conditions. The Twin has a horizontal resolution of about 5 km, which is extremely valuable from a scientific point of view, but also computationally very expensive. The project presented here is led by Predictia together with BSC, CSIC and FZJ, with the objective of developing an ML-based Climate Emulator of the Twin. This Climate Emulator aims to replicate key aspects of the existing Climate Digital Twin, using a fraction of the computational resources required by the full simulation.

Second presentation: "ML-based temporal downscaler for data compression" by Alberto Sanchez-Marroquin

Beyond being computationally expensive to produce, the Climate Digital Twin produced by DestinE also requires a massive amount of data storage. As of September 2026, the project has generated a portfolio of about 9 PB, a volume prohibitive to hold on HPC disk in full, as well as difficult to use. Therefore this project is developing an ML-based application to compress Climate Digital Twin datasets, in order to help the scientific community and stakeholders work with such a large dataset. The compression application is a temporal downscaler model based on a Graph Neural Network which operates similarly to existing ML-based forecasts. The key difference is that, instead of generating an atmospheric trajectory from a few temporal initial states, our model receives both an initial and a final atmospheric state and reconstructs the trajectory between them. Once trained, the model enables compression: only a sparse set of timestamps (for instance, one every 24 hours) needs to be stored. The intermediate time steps can then be regenerated on demand from these stored states using the temporal downscaler. This reduces storage requirements substantially, at the cost of a small reconstruction error. Additionally, this method of compression can be combined with traditional compression methods to achieve even higher compression ratios.