Correcting weather and climate models by machine learning nudged historical simulations

Abstract

Existing ML-based atmospheric models are not suitable for climate prediction, which requires long-term stability and physical consistency. We present ACE (AI2 Climate Emulator), a 200M-parameter, autoregressive machine learning emulator of an existing comprehensive 100-km resolution global atmospheric model. The formulation of ACE allows evaluation of physical laws such as the conservation of mass and moisture. The emulator is stable for 10 years, nearly conserves column moisture without explicit constraints and faithfully reproduces the reference model's climate, outperforming a challenging baseline on over 80% of tracked variables. ACE requires nearly 100x less wall clock time and is 100x more energy efficient than the reference model using typically available resources.

Assessing the Potential of Deep Learning for Emulating Cloud Superparameterization in Climate Models With Real‐Geography Boundary Conditions

“…In hindsight, it would be logical to complement the benefits of hyperparameter tuning with such constraints-an important topic for future work. It is also possible that skill in the precipitation field would benefit from enforcing a consistency between it and the column moistening that is better emulated, as in Beucler et al (2019) or Watt-Meyer et al (2021).…”

“…To cite a few, RFs with deep trees quickly become computationally expensive for large data sets, requiring large storage capacity which could prevent taking full advantage of Graphics Processing Unit (GPU) infrastructure (Yuval & O'Gorman, 2020b). RFs may struggle to capture local patterns in the atmosphere as well (Watt-Meyer et al, 2021). For these reasons, we leave RFs for future work.…”

“…Indeed, RFs have some advantages, including automatically respecting physical constraints that are linear in their outputs, such as energy conservation, and positive-definite precipitation, which are not guaranteed in DNNs (Yuval & O'Gorman, 2020b). Furthermore, Watt-Meyer et al (2021) have demonstrated that RFs can be used to correct, or "nudge" parameterizations and reduce simulation errors even for realistic convection. However, there are also ways to enforce such constraints in DNNs (Beucler et al, 2019;Kumar et al, 2020), and RFs also come with disadvantages.…”

fv3gfs-wrapper: a Python wrapper of the FV3GFS atmospheric model

“…The separation of physics and dynamics steps in the code makes it clear that the machine learning update is applied at the end of a physics step and is included in any intermediate restart data. The random forest model used in this example is trained according to the approach in Watt-Meyer et al (2021), with a small number of trees and layers chosen to decrease model size. As a proof of concept, the example model has not been tuned for stability and may crash if run for longer than 6 h or if a run directory other than the example provided is used.…”

“…1 Introduction FV3GFS (Finite-Volume Cubed-Sphere Global Forecast System) (Zhou et al, 2019) is a prototype of the operational Global Forecast System of the National Centers for Environmental Prediction. In this document when we say FV3GFS we are referring specifically to the atmospheric component of the US National Oceanic and Atmospheric Administration (NOAA) Unified Forecast System (UFS; https: //ufscommunity.org/, last access: 21 May 2021) for operational numerical weather prediction.…”