- 1Stanford University, Earth system science, United States of America (aditi_sheshadri@stanford.edu)
- 2University of Oxford, Oxford, United Kingdom
Atmospheric gravity waves (GWs) present a challenge to climate prediction since most of their spectrum is not resolved in global climate models and good observational constraints on GW activity do not exist. One of the long-standing approximations made in gravity wave parameterizations (GWPs) is the assumption of purely vertical propagation of these waves (no horizontal nonlocality). I will present recent developments in my group on using machine learning (ML) methods to aid the parameterization of the effects of breaking atmospheric gravity waves in global climate models. These efforts have advanced through two distinct approaches: a) replacing existing physics-based GW parameterizations with ML algorithms, and b) using ML methods to aid in the calibration of existing physics-based parameterizations. On a) I will describe ML-based GW parameterizations that incude various degrees of nonlocality, including a globally nonlocal scheme that is trained on high-resolution data. On b), I will present results on using methods including Ensemble Kalman methods and Bayesian methods to calibrate parameters in physics-based GWPs, as well as to estimate parametric uncertainty in climate projections.
How to cite: Sheshadri, A., Gupta, A., King, R., and Mansfield, L.: Using ML-based methods to improve the representation of atmospheric gravity waves in climate models, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13526, https://doi.org/10.5194/egusphere-egu25-13526, 2025.