CAM-NET: An AI Model for Whole Atmosphere with Thermosphere and Ionosphere Extension

Jiahui Hu, Wenjun Dong · 2025

Read CAM-NET: An AI Model for Whole Atmosphere with Thermosphere and Ionosphere Extension by Jiahui Hu, Wenjun Dong (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-19340.txt.

Introduction

Jiahui Hu and Wenjun Dong build a fast stand-in for a whole-atmosphere physics model, in a 2025 paper. The printed title, naming a geometry-aware emulator from surface to ionosphere, differs from the shorter catalog title.

The reference model, WACCM-X, simulates air from the ground to the upper atmosphere but costs a fortune to run. CAM-NET, a neural operator built for spheres, learns from its three-hourly output. It predicts winds, temperature, electron density, and ion drift. A light add-on module extends it to plasma variables. On held-out runs it kept the large patterns and stayed stable for days. Fine wavy structures near 90 kilometers were damped, an honest weakness. Cheap ensembles are the prize.

Catalog record

CREATORJiahui Hu, Wenjun Dong
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE28 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-19340

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

CAM-NET: An AI Model for Whole Atmosphere with Thermosphere and Ionosphere Extension Jiahui Hu, Wenjun Dong Abstract Abstract We present Compressible Atmospheric Model-Network (CAM-NET), an AI model designed to predict neutral atmospheric variables from the Earth’s surface to the ionosphere with high accuracy and computational efficiency. Accurate modeling of the entire atmosphere is critical for understanding the upward propagation of gravity waves, which influence upper-atmospheric dynamics and coupling across atmospheric layers. CAM-NET leverages the Spherical Fourier Neural Operator (SFNO) to capture global-scale atmospheric dynamics while preserving the Earth’s spherical structure. Trained on a decade of datasets from the Whole Atmosphere Community Climate Model with thermosphere and ionosphere eXtension (WACCM-X), CAM-NET demonstrates accuracy comparable to WACCM-X while achieving a speedup of over 1000x in inference time, can provide one year simulation within a few minutes once trained. The model effectively predicts key atmospheric parameters, including zonal and meridional winds, temperature, and time rate of pressure. Inspired by traditional modeling approaches that use external couplers to simulate tracer transport, CAM-NET introduces a modular architecture that explicitly separates tracer prediction from core dynamics. The core backbone of CAM-NET focuses on forecasting primary physical variables (e.g., temperature, wind velocity), while tracer variables are predicted through a lightweight, fine-tuned model. This design allows for efficient adaptation to specific tracer scenarios with minimal computational cost, avoiding the need to retrain the entire model. We have validated this approach on the O^{2} tracer, demonstrating strong performance and generalization capabilities. Plain Language A new AI model that predicts atmospheric conditions from the Earth’s surface to the ionosphere with high accuracy and speed. It uses a specialized neural network, based on a geometry-aware neural operator, to recognize global weather patterns while keeping the shape of Earth consistent. Trained on ten years of climate data, CAM-NET matches the accuracy of existing climate models but works much faster. Large models that aren’t reusable across applications miss the advantage of their low inference cost. To address this, we enhanced CAM-NET by integrating a tracer tracking module. CAM-NET combines a backbone with lightweight fine-tuning models for tracer prediction. Similar to traditional models that use external couplers to simulate tracers, CAM-NET separates tracer prediction from core dynamics. The backbone predicts core physical variables (e.g., temperature, velocity), while tracers are predicted separately. This setup enables low-cost adaptation to specific scenarios without retraining the full model. As such, CAM-NET serves as a complementary tool to traditional numerical models, offering a flexible and cost-effective solution for simulating the full atmospheric state. 1. Introduction Improving atmospheric modeling and forecasting are high priorities for many broad-ranging research interests. Physics-based numerical weather prediction (NWP) models project future weather conditions by numerically solving the differential equations that describe complex physical interactions between the atmosphere, land, and ocean systems. European Center for Medium-Range Weather Forecasts (ECMRWF) develops a general circulation model, the Integrated Forecasting System (IFS), to generate forecasts with height ranging from Earth’s surface to the atmosphere, and with a time span from several days to one year Roberts et al. (2018) . Whole Atmosphere Community Climate Model with Thermosphere and Ionosphere eXtension (WACCM-X) is another ground-to-topside NWP, developed by National Center Atmospheric Research (NCAR), which specifically solves for the comprehensive interactions between atmosphere and ionosphere H-L. Liu et al. (2018) . However, NWP models often generate deterministic outputs that are solely dependent on initial conditions, without explicitly incorporating the range of uncertainties inherent in the evolution of weather systems Bülte et al. (2025) . This deterministic nature limits the model’s ability to convey the probabilistic distributions of states. In addition, the underlying physics of atmospheric processes are governed by highly nonlinear and high-dimensional partial differential equations. Solving these equations accurately requires immense computational resources and precise initialization Ren et al. (2021) . These factors not only constrain the accuracy of long-term weather forecasts but also hinder the feasibility of real-time applications that demand both speed and reliability. To address these challenges, researchers have increasingly turned to data-driven models that leverage deep learning to enhance the forecasting capabilities of NWP models. Notable examples Bülte et al. (2025) , such as FourCastNet Pathak et al. (2022) , GraphCast Lam et al. (2023) , Pangu-Weather Bi et al. (2023) , Fuxi L. Chen et al. (2023) ; Zhong et al. (2024) , Fengwu K. Chen et al. (2023) , ClimaX Nguyen et al. (2023) , GenCast Price et al. (2023) , are trained on ERA5 datasets Hersbach et al. (2020) , the fifth generation of datasets produced by ECMRWF. The model outputs have demonstrated that AI models can match or even exceed the forecast accuracy of these operational weather models in medium-range forecasts, while significantly reducing inference time and computational costs. One of the models, FourCastNet, is proven to be a powerful method in solving long-range dependencies in spatio-temporal data Pathak et al. (2022) . FourCastNet is built based on Adaptive Fourier Neural Operator (AFNO) learning Li et al. (2020) , which extends the Fourier Neural Operator (FNO) framework by mixing tokens through discrete Fourier transforms (DFT), enabling efficient modeling of long-range dependencies Guibas et al. (2021) . However, DFTs introduce visual and spectral artifacts, along with significant dissipation, when learning operators in spherical coordinates because they incorrectly assume a flat underlying geometry for the Earth Bonev et al. (2023) . Moreover, neural operators are prone to spectral bias, reducing their sensitivity to high-frequency modes Khodakarami et al. (2025) . As a result, they struggle to represent the fine-scale atmospheric dynamics (i.e. gravity waves) in multiscale physical systems. The models mentioned above are primarily designed for predictions from the Earth surface to the stratosphere. However, the atmosphere exhibits distinct dynamical regimes across altitudes from surface to thermosphere, due to change in density, composition, and dominant forcing mechanisms H-L. Liu et al. (2024) . Mesoscale structures such as Gravity Waves (GWs), originate primarily from tropospheric source such as deep convection and topography, propagate vertically and eventually break due to increasing molecular viscosity and background wind shears Fritts & Alexander (2003) . Wave dissipation generates secondary GWs with broader spatial scales and faster phase speeds, and some of which can reach the thermosphere, where they modulate neutral density and ionospheric plasma Fritts & Lund (2011) ; Vadas & Liu (2009) ; Dong et al. (2020) ; Fritts et al. (2020) . Observational evidence, such as electron perturbations, has been linked to these lower atmospheric layers Forbes et al. (2016) ; Azeem et al. (2015) ; H. Liu et al. (2017) . Therefore, a unified whole atmosphere model spanning from the surface to thermosphere is essential for accurately capturing these interactive processes. We introduce CAM-NET , a neural architecture initially developed based on the Adaptive Fourier Neural Operator (AFNO) framework Dong et al. (2023) . In this work, we also present an alternative built on the Spherical Fourier Neural Operator (SFNO) Bonev et al. (2023) , which is specifically tailored to Earth’s spherical geometry. SFNO enables the capture of global-scale, whole-atmosphere dependencies with high spectral accuracy, making it well-suited for climate and weather modeling tasks. To quantify vertical coupling processes, tracer variables play a critical role Keneshea et al. (1979) . Among them, molecular oxygen ( \mathrm{O}_{2} ) serves as an ideal tracer due to its

Read the complete text