Larson et al. (2025) QuadTune version 1: a regional tuner for global atmospheric models

Identification

• Journal: Geoscientific model development
• Year: 2025
• Date: 2025-12-09
• Authors: Vincent E. Larson, Zhun Guo, Benjamin A. Stephens, Colin M. Zarzycki, Gerhard Dikta, Yun Qian, Shaocheng Xie
• DOI: 10.5194/gmd-18-9767-2025

Research Groups

• University of Wisconsin – Milwaukee, Milwaukee, WI, USA
• National Key Laboratory of Earth System Numerical Modeling and Application, Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing, China
• National Center for Atmospheric Research, Boulder, CO, USA
• Department of Meteorology and Atmospheric Science, The Pennsylvania State University, University Park, PA, USA
• Department of Medical Engineering and Technomathematics, Fachhochschule Aachen, Jülich, Germany
• Pacific Northwest National Laboratory, Richland, WA, USA
• Lawrence Livermore National Laboratory, Livermore, CA, USA

Short Summary

QuadTune is a novel, inexpensive regional tuner for global atmospheric models that uses an uncorrelated quadratic emulator to quickly reduce parametric errors. In an example application, it demonstrated a reduction in Shortwave Cloud Radiative Forcing (SWCF) root mean square error (RMSE) from 12.4 W m⁻² to 10.4 W m⁻².

Objective

• To introduce and document QuadTune, an inexpensive regional tuner for global atmospheric models, designed to quickly retune models after structural changes and provide diagnostic insights into remaining structural errors and tuning trade-offs.

Study Configuration

• Spatial Scale: Global atmospheric model (EAM) divided into 20° longitude by 20° latitude regions (tiles) for tuning, with a sensitivity analysis also using 30° longitude by 30° latitude regions.
• Temporal Scale: Tuning process based on 1-year (or 2-year for sensitivity analysis) global sensitivity simulations; final evaluation of tuned model performed over 5-year simulations.

Methodology and Data

• Models used:
  • Global Atmospheric Model: Energy Exascale Earth System Model (E3SM) atmosphere component (EAM), a development version close to EAMv3.0.0.
  • Parameterization: Cloud Layers Unified By Binormals (CLUBB), specifically the "taus" code option.
  • Tuning Method: QuadTune, which employs an uncorrelated quadratic emulator (response surface) and minimizes a least-squares loss function using the Powell method.
• Data sources:
  • Observations for tuning: Shortwave Cloud Radiative Forcing (SWCF) from version 4.1 of the Clouds and the Earth's Radiant Energy System (CERES) Energy Balanced And Filled (EBAF) dataset.

Main Results

• QuadTune reduced the RMSE of SWCF from 12.4 W m⁻² (untuned default EAM-taus) to 10.4 W m⁻² (quad-tuned), closely approaching the 10.1 W m⁻² achieved by laborious hand-tuning.
• The global-mean bias of the quad-tuned model (0.32 W m⁻²) was better than that of the hand-tuned model (1.85 W m⁻²).
• Nonlinear parameter dependence was found to be significant, with quadratic contributions to sensitivity being sizable across global regions, indicating that a single linear sensitivity is often insufficient.
• QuadTune prioritizes bias reduction in regions with large initial biases and high sensitivities (e.g., stratocumulus regions), sometimes at the expense of other regions.
• The study identified tuning trade-offs (e.g., improving one region worsens another) and stubborn biases (regions with non-negligible bias but low sensitivity to parameters, such as the Canadian Arctic region 1_14).
• The most influential parameters for tuning were n2_thresh, c8, sfc, and n2, with c8 being crucial for restoring global radiative balance despite its low correlation with local bias patterns.
• The quality of tuning results showed only modest sensitivity to the size of tuning regions (20°x20° vs 30°x30°) and the duration of sensitivity simulations (1 year vs 2 years).
• Including a single parameter interaction term in the emulator improved the RMSE of SWCF from 10.4 W m⁻² to 10.1 W m⁻².

Contributions

• Introduction of QuadTune, an inexpensive and efficient regional tuner for global atmospheric models that significantly reduces the number of required global simulations (2P+1).
• Development of a suite of diagnostic plots and a quasi-linear sensitivity matrix (S⁺) to enhance explainability and visualize structural model errors, such as stubborn biases and tuning trade-offs.
• Demonstration of QuadTune's practical utility for quickly retuning models after structural modifications (e.g., switching CLUBB parameterization options).
• Highlighting the importance of nonlinear parameter dependence in atmospheric model tuning and providing a framework for its analysis.

Funding

• US Department of Energy (grant nos. DE-AC52-07NA27344 and DE-SC0025252)
• National Natural Science Foundation of China (grant no. 42175164)
• Basic Scientific Research Project of Institute of Atmospheric Physics during the 14th Five-year Plan period
• Climate Process Team (CPT) under Grant AGS-1916689 from the National Science Foundation (NSF)
• Grant NA19OAR4310363 from the National Oceanic and Atmospheric Administration (NOAA)

Citation

@article{Larson2025QuadTune,
  author = {Larson, Vincent E. and Guo, Zhun and Stephens, Benjamin A. and Zarzycki, Colin M. and Dikta, Gerhard and Qian, Yun and Xie, Shaocheng},
  title = {QuadTune version 1: a regional tuner for global atmospheric models},
  journal = {Geoscientific model development},
  year = {2025},
  doi = {10.5194/gmd-18-9767-2025},
  url = {https://doi.org/10.5194/gmd-18-9767-2025}
}