Editors’ Highlights are summaries of recent papers by AGU journal editors.
sauce: AGU progress
Machine learning (ML)-based models have great potential to enhance and potentially transform simulations of Earth’s weather and climate over synoptic, seasonal, annual, and multidecadal time scales. However, ML-based models must also produce results that are consistent with the physical laws of the Earth system. Although ML-based models have been tested for weather forecasting, their ability to produce reasonable responses in long-term simulations under relevant forcing over weather-climate timescales remains uncertain. Therefore, it is essential to perform extensive evaluations over different timescales. Additionally, it is important to understand the extent to which new ML techniques can complement traditional physics-based models.
Chen et al. [2026] We perform a series of tests that target systems that are subject to aggregate scale, annual scale, and long-term deallocation enforcement. This study uses a hybrid model called NeuralGCM, which combines traditional Earth system modeling and ML approaches. In a set of idealized experiments, NeuralGCM produces performance similar to traditional physically-based Earth system models. However, we found several limitations in simulating extratropical cyclone strength, atmospheric wave response, and stratospheric warming and circulation responses. In general, the combination of established physically-based modeling and ML represents a promising pathway to achieve weather and climate analysis that requires less computational time.
Citation: Chen, Z., Leung, LR, Zhou, W., Lu, J., Lubis, SW, Liu, Y., et al. (2026). Hierarchical testing of hybrid machine learning and physics global atmospheric models. AGU progress7, e2025AV002075. https://doi.org/10.1029/2025AV002075
—Don Webles, editor AGU progress
