Stable machine-learning parameterization of subgrid processes for climate modeling at a range of resolutions (2001.03151v2)

Abstract: Global climate models represent small-scale processes such as clouds and convection using quasi-empirical models known as parameterizations, and these parameterizations are a leading cause of uncertainty in climate projections. A promising alternative approach is to use machine learning to build new parameterizations directly from high-resolution model output. However, parameterizations learned from three-dimensional model output have not yet been successfully used for simulations of climate. Here we use a random forest to learn a parameterization of subgrid processes from output of a three-dimensional high-resolution atmospheric model. Integrating this parameterization into the atmospheric model leads to stable simulations at coarse resolution that replicate the climate of the high-resolution simulation. The parameterization obeys physical constraints and captures important statistics such as precipitation extremes. The ability to learn from a fully three-dimensional simulation presents an opportunity for learning parameterizations from the wide range of global high-resolution simulations that are now emerging.

Stable Machine-Learning Parameterization of Subgrid Processes for Climate Modeling

The research presented in this paper proposes a novel approach to enhancing climate models by integrating machine learning-based parameterizations of subgrid processes. These subgrid processes, such as convection and cloud formation, are typically parameterized using heuristic models due to limited computational capabilities, causing considerable uncertainty in global climate projections.

This research utilizes a Random Forest (RF) machine learning model to derive parameterizations from high-resolution three-dimensional atmospheric simulations. The paper effectively demonstrates the capability of these RF parameterizations to simulate stable climate scenarios at a coarse resolution while preserving essential statistical properties and physical constraints, particularly in reproducing the climates of high-resolution simulations.

Key Contributions

1. High-resolution Training: The RF model is trained on outputs from the System for Atmospheric Modeling, a high-resolution atmospheric model that captures processes across diverse spatiotemporal scales, including convection and large-scale atmospheric circulations.
2. Implementation of RF Parameterization: Through rigorous parameterization, the RF model is integrated into a coarser atmospheric model grid while ensuring physical constraints such as energy conservation and non-negativity of surface precipitation. The RF captures both the thermodynamic and moisture dynamics, gaining substantial performance improvement over conventional parameterizations.
3. Scale-awareness: One innovative aspect of this work is its ability to generalize the RF parameterization across different grid-spacing scales, a challenge that traditional parameterizations face due to their reliance on scale-dependent assumptions. The model operates effectively even in the so-called "gray-zone" of model resolution, where grid spacings are not sufficiently fine to resolve convective processes directly.

Significant Results

• The RF parameterization showed impressive stability across long-duration climate simulations, replicating key climate statistics from the high-resolution simulations, such as precipitation extremes, mean precipitation distributions, and zonal circulations.
• The RF model demonstrated superior offline and online performance, particularly at smaller grid spacings, addressing a vital consideration in atmospheric modeling—the need for scale-aware parameterizations that adapt to the increasing resolution of global climate models.
• The research indicates a significant reduction in computational resources, by a factor of 30 for an x8 coarse grid resolution, suggesting potential for application in extensive climate modeling studies feasibly.

Implications and Future Directions

The use of an RF model highlights the growing utilization of machine learning for complex environmental simulations. The parameterization approach respects physical principles while providing flexibility across spatial resolutions, an essential step forward in the quest to improve model predictability and reliability.

Potential future work includes the extension of these methods to land regions and more complex geographic domains, which this paper did not address. Furthermore, it could be valuable to explore the interplay between RF parameterizations and neural networks, which may offer advantages in memory use and potentially enhance online performance in diverse atmospheric conditions.

The findings hold substantial promise for more robust climate predictions, aiding climate researchers and policymakers in understanding and mitigating the impacts of climate variability and change. The integration of machine learning techniques into traditional climate model frameworks signifies an evolving research frontier in computational climate science.