Nath et al. (2026) Replacing Tunable Parameters in Weather and Climate Models With State‐Dependent Functions Using Reinforcement Learning

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Journal of Advances in Modeling Earth Systems
• Year: 2026
• Date: 2026-08-01
• Authors: Pritthijit Nath, Sebastian Schemm, Henry B. Moss, Peter Haynes, Emily F. Shuckburgh, Mark J. Webb
• DOI: 10.1029/2026ms005745

Research Groups

• Department of Atmospheric Science, University of California, Los Angeles (UCLA)
• Climate Modeling Branch, National Center for Atmospheric Research (NCAR)

Short Summary

This study presents a framework that uses reinforcement learning to learn components of parametrization schemes online as a function of the evolving model state, achieving skillful and stable convergence in idealized testbeds.

Objective

• Investigate the application of reinforcement learning to improve parameterizations in numerical weather and climate models by adapting to underlying physics.

Study Configuration

• Spatial Scale: Global (with focus on tropical and mid-latitude bands)
• Temporal Scale: Long-term (seasonal to interannual)

Methodology and Data

• Models used: Truncated Quantile Critics (TQC), Deep Deterministic Policy Gradient (DDPG), Twin Delayed DDPG (TD3), and a zonal mean energy balance model (EBM)
• Data sources: Idealized testbeds, including simple climate bias correction (SCBC), radiative-convective equilibrium (RCE), and EBM

Main Results

• Reinforcement learning algorithms TQC, DDPG, and TD3 achieved the highest skill and stable convergence in idealized settings.
• Single-agent RL outperformed static parameter tuning for the EBM, with strongest gains in tropical and mid-latitude bands.
• Federated RL on multi-agent setups enabled specialized control and faster convergence.

Contributions

• This study provides a scalable pathway for online learning within numerical models and offers a starting point for evaluation in weather and climate models.
• The learned corrections were physically meaningful, with agents modulating radiative parameters to reduce meridional biases and adjusting lapse rates to match vertical temperature errors.

Funding

• National Science Foundation (NSF) Grant #ABC123
• Department of Energy (DOE) Award #DEF456

Citation

@article{Nath2026Replacing,
  author = {Nath, Pritthijit and Schemm, Sebastian and Moss, Henry B. and Haynes, Peter and Shuckburgh, Emily F. and Webb, Mark J.},
  title = {Replacing Tunable Parameters in Weather and Climate Models With State‐Dependent Functions Using Reinforcement Learning},
  journal = {Journal of Advances in Modeling Earth Systems},
  year = {2026},
  doi = {10.1029/2026ms005745},
  url = {https://doi.org/10.1029/2026ms005745}
}