PseudospectralNet: Toward Hybrid Atmospheric Models for Climate Simulations

Journal article

Plain Language Summary: Many machine learning (ML) models that aim to learn atmospheric dynamics are not suitable for climate simulations because long simulations often suffer from numerical inaccuracies and instabilities. In this project, we investigate whether adding physics‐based knowledge to ML models helps to improve those long simulations and also make forecasts of the models more accurate. For this purpose we divide a simple atmospheric model into its individual components and use their outputs at every time step of the simulation as an input to artificial neural networks that learn how to forecast example data accurately with the information from the data itself and the physics‐based components. We find that such a hybrid model outperforms purely data‐driven ML methods in our applications.

Authors: Gelbrecht, M; Klöwer, M; Boers, N

• Publication status: Published
• Peer review status: Peer reviewed
• Publisher: Wiley
• Journal: Journal of Advances in Modeling Earth Systems
• Volume: 17
• Issue: 10
• Article number: e2025MS004969
• Publication date: 2025-10-18
• Acceptance date: 2025-09-15
• DOI: 10.1029/2025ms004969
• EISSN: 1942-2466
• ISSN: 1942-2466
• Language: English
• Keywords: scientific machine learning; hybrid model; machine learning; atmospheric model; quasigeostrophic model; differentiable programming