Aiming for a learnable climate model

Artificial intelligence (AI) is bringing significant changes to atmospheric science, especially with the introduction of large-scale AI weather models such as Pangu-Weather and GraphCast. However, parallel to these advances are questions about the consistency of these models with fundamental physical principles.

Previous studies have demonstrated that Pangu-Weather can accurately reproduce certain climate patterns such as tropical Gill responses and extratropical teleconnections through qualitative analysis. However, quantitative research has revealed that there are significant differences in wind components such as divergent winds and strata winds within current AI weather models. Despite these findings, there are still concerns that the importance of physics in climate science is sometimes overlooked.

“Qualitative evaluations showed that AI models were able to understand and learn from spatial patterns in weather and climate data, while quantitative approaches highlighted limitations. It struggles to learn patterns and instead relies solely on total wind speed,'' explains Professor Gang Huang of the Institute of Atmospheric Physics (IAP), Chinese Academy of Sciences.

“This highlights the need for comprehensive dynamic diagnostics of AI models. Only through holistic analysis can we deepen our understanding and impose the necessary physical constraints.”

Researchers, including collaborators from IAP, Seoul National University, and Tongji University, are moving beyond the concept of “either-or” scenarios and advocating a joint approach of AI and physics in climate modeling.

Professor Huang said, “Although AI is good at capturing spatial relationships within weather and climate data, it struggles with subtle physical factors such as divergent winds and strata winds. “This highlights the need for rigorous dynamic diagnostics to enforce this.”

was announced on Advances in atmospheric scienceTheir perspective paper shows how to impose both soft and hard physical constraints on AI models to ensure consistency with known atmospheric dynamics.

Furthermore, the team advocates a transition of the parameterization scheme from offline to online to achieve global optimality in model weighting, thereby achieving a balanced combination of fully coupled physics and AI. Facilitate climate modeling. “This integration enables iterative optimization, transforming the model into a truly learnable system,” said Dr. Ya Wang.

Researchers recognize the importance of community collaboration and promote a culture of openness, comparability, and reproducibility (OCR). They believe that by adopting principles similar to those in the AI ​​and computer science communities, they will foster a culture that is conducive to developing truly learnable climate models.

In summary, by integrating the spatial capabilities of AI with the fundamental principles of physics and fostering a collaborative community, researchers aim to achieve climate models that seamlessly blend AI and physics. It represents an important advance in climate science.