2026-02-04 スタンフォード大学
Gravity waves can produce parallel bands of clouds, as seen in this 2013 view of Lake Superior from the International Space Station. | ISS Crew Earth Observations experiment and Image Science & Analysis Laboratory, Johnson Space Center
＜関連情報＞
- https://news.stanford.edu/stories/2026/02/atmospheric-gravity-waves-climate-modeling
- https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2025MS005075
- https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2025MS004977
AI基盤モデルの微調整によるサブグリッドスケールパラメータ化の開発：大気重力波のケーススタディ Finetuning AI Foundation Models to Develop Subgrid-Scale Parameterizations: A Case Study on Atmospheric Gravity Waves
Aman Gupta, Aditi Sheshadri, Sujit Roy, Johannes Schmude, Vishal Gaur, Wei Ji Leong, Manil Maskey, Rahul Ramachandran
Journal of Advances in Modeling Earth Systems Published: 15 November 2025
DOI:https://doi.org/10.1029/2025MS005075
Abstract
Global climate models parameterize a range of atmospheric-oceanic processes, including gravity waves (GWs), clouds, moist convection, and turbulence, that cannot be sufficiently resolved. These subgrid-scale closures for unresolved processes are a substantial source of model uncertainty. Here, we present a new approach to developing machine learning (ML) parameterizations of small-scale climate processes by fine-tuning a pre-trained AI foundation model (FM). FMs are largely unexplored in climate research. A pre-trained encoder-decoder from a 2.3 billion parameter FM (NASA and IBM Research’s Prithvi WxC)—which contains a latent probabilistic representation of atmospheric evolution—is fine-tuned (or reused) to create a deep learning parameterization for atmospheric gravity waves (GWs); a process unseen during pre-training. The parameterization captures GW effects for a coarse-resolution climate model by learning the fluxes from an atmospheric reanalysis with 10 times finer resolution. A comparison of monthly averages and instantaneous evolution with a machine learning model baseline (an Attention U-Net) reveals superior predictive performance of the FM parameterization throughout the atmosphere, even in regions excluded during pre-training. This performance boost is quantified using the Hellinger distance, which is 0.11 for the baseline and 0.06 for the fine-tuned model. Our findings emphasize the versatility and reusability of FMs, which could be used to accomplish a range of atmosphere- and climate-related applications, leading the way for the creation of observations-driven and physically accurate parameterizations for more earth system processes.
Plain Language Summary
Climate models struggle to accurately capture the physical effects of small-scale atmospheric processes like gravity waves, turbulence, and clouds, which are critical to accurately predicting future climate states. These processes evolve on scales finer than typical model grid resolutions. As a result, they continue to rely on approximations, known as physical parameterizations, to represent their missing effects. The use of parameterizations introduces uncertainty and makes climate predictions less reliable. Here, we propose a new approach to improving these parameterizations using modern advances in deep learning. Specifically, we use Prithvi WxC, a large AI model trained on multiple decades of one reanalysis, and fine-tune it using limited years of gravity wave (GW) data from another reanalysis to develop an emulator capable of predicting a physically consistent atmospheric GW flux evolution. The novel approach of leveraging a large AI model pre-trained on vast volumes of atmospheric data and augmenting it with limited process-specific data allows the creation of compact and easily trainable data-driven physical parameterizations. While we focus on gravity waves, our approach is flexible and can be generalized to developing data-driven parameterizations of other earth system processes.
Key Points
- AI weather foundation models can be fine-tuned to create climate model parameterizations for subgrid-scale processes like gravity waves
- The fine-tuned machine learning (ML) parameterization beats existing ML benchmarks, predicting accurate wave fluxes and variability
- Despite predicting accurate monthly averages and strong wave events, ML models continue to struggle with the prediction of small flux values
大気重力波の気候モデル表現のための非局所深層学習パラメータ化のオフラインパフォーマンス Offline Performance of a Nonlocal Deep Learning Parameterization for Climate Model Representation of Atmospheric Gravity Waves
Aman Gupta, Aditi Sheshadri, Sujit Roy, Valentine Anantharaj
Journal of Advances in Modeling Earth Systems Published: 22 October 2025
DOI:https://doi.org/10.1029/2025MS004977
Abstract
Gravity waves (GWs) make crucial contributions to the middle atmospheric circulation. Yet, their climate model representation remains inaccurate, leading to key circulation biases. This study introduces a set of three neural networks (NNs) that learn to predict GW fluxes (GWFs) from multiple years of high-resolution ERA5 reanalysis. The three NNs: a <?XML:NAMESPACE PREFIX = “[default] http://www.w3.org/1998/Math/MathML” NS = “http://www.w3.org/1998/Math/MathML” /> ANN, a ANN-CNN, and an Attention UNet embed different levels of horizontal nonlocality in their architecture and are capable of representing nonlocal GW effects that are missing from current operational GW parameterizations. The NNs are evaluated offline on both time-averaged statistics and time-evolving flux variability. All NNs, especially the Attention UNet, accurately recreate the global GWF distribution in both the troposphere and the stratosphere. Moreover, the Attention UNet most skillfully predicts the transient evolution of GWFs over prominent orographic and nonorographic hotspots, with the model being a close second. Since even ERA5 does not resolve a substantial portion of GWFs, this deficiency is compensated by subsequently applying transfer learning on the ERA5-trained ML models for GWFs from a 1.4 km global climate model. It is found that the re-trained models both (a) preserve their learning from ERA5, and (b) learn to appropriately scale the predicted fluxes to account for ERA5’s limited resolution. Our results highlight the importance of embedding nonlocal information for a more accurate GWF prediction and establish strategies to complement abundant reanalysis data with limited high-resolution data to develop machine learning-driven parameterizations for missing mesoscale processes in climate models.
Plain Language Summary
Gravity waves (GWs) are ubiquitous atmospheric oscillations generated by flow disturbances in the atmosphere. Since they operate on smaller scales than a climate model can resolve, their effects are mostly unresolved in coarse-resolution climate models. So, climate models typically parameterize/approximate their effects, but these parameterizations can often be oversimplified, leading to physical inaccuracies in models. We propose a set of three fully machine learning (ML)-based parameterizations whose architectures are chosen to capture both horizontal and vertical GW propagation: single column, multiple columns, and globally nonlocal, to learn GW effects from data. Following training on multiple (four) years of modern reanalysis, these purely data-driven schemes generate accurate flux statistics, time evolution, and variability. The globally nonlocal ML model offers the best prediction, indicating the importance of nonlocality for data-driven GW schemes. We subsequently re-train the models on 4 months of a 1.4 km climate model and find that iteratively training on high-volume, low-resolution reanalysis and low-volume, high-resolution climate model output allows the model to learn GW effects from both data sets effectively. Our results establish the capability of ML-based schemes to learn essential GW physics from a mix of data, to represent these missing effects in climate models, and improve their prediction.
Key Points
- Three machine learning schemes are trained on global, resolved wave fluxes, embedding three different levels of horizontal nonlocality
- The three neural networks accurately reproduce both time-averaged statistics and transient flux variability over prominent gravity wave hotspots
- Transfer learning on a 1.4 km climate model improves flux prediction and variability around the tropical quasi-biennial oscillation and Antarctic final warming
