Implicit learning of convective organization explains precipitation stochasticity

Sara Shamekh, Kara D. Lamb, Y. Huang, Pierre Gentine

TL;DR

Climate models struggle to accurately predict precipitation intensity, especially extreme events, because they miss crucial subgrid-scale cloud organization. This paper uses machine learning to implicitly learn this "organization" from high-resolution simulations, significantly improving precipitation prediction accuracy and capturing extreme events.

What Problem Does This Paper Address?

Accurate precipitation prediction is vital for understanding climate change impacts, yet current climate models (like General Circulation Models or GCMs) fall short, particularly with extreme precipitation events. A key reason for this deficiency is that GCMs operate at a coarse resolution (around 100 km), which is too large to explicitly resolve the small-scale processes of convection and cloud organization. Instead, these processes are approximated through "parameterizations." Traditional parameterizations often treat convective clouds as randomly distributed, neglecting how their clustering and interaction (organization) can significantly influence moisture distribution, cloud development, and ultimately, precipitation intensity and variability. This leads to models that tend to predict too much light rain and underestimate extremes. The paper aims to bridge this gap by investigating if learning subgrid-scale organization can improve precipitation prediction in coarse-resolution models.

Methods

The researchers employed a machine learning approach, specifically neural networks (NNs), to tackle this problem. They developed two main NN architectures:

  1. Baseline-NN: This is a standard feed-forward neural network that takes coarse-scale atmospheric variables (like precipitable water, sea surface temperature, near-surface humidity, and temperature) as input and predicts coarse-scale precipitation. This serves as a benchmark to assess how well precipitation can be predicted using only large-scale information.
  2. Org-NN: This more advanced architecture combines the baseline-NN with an autoencoder. The autoencoder is trained on high-resolution precipitable water fields to learn a compressed, low-dimensional representation of subgrid-scale organization, which they call the "org" metric. This "org" metric is then fed as an additional input to the precipitation prediction NN, alongside the coarse-scale variables. The autoencoder and the precipitation predictor are trained simultaneously, allowing the autoencoder to learn "org" features that are most relevant for improving precipitation prediction.

To predict the "org" metric itself for use in GCMs (where high-resolution data isn't available), they also developed a separate NN that learns to predict "org" from historical coarse-scale variables and/or previous "org" values, demonstrating its temporal memory.

Datasets

The primary dataset used is from the System for Atmospheric Modeling (SAM), specifically from the DYAMOND Phase 2 Intercomparison Project. This project generated global, high-resolution (4.2 km) storm-resolving simulations of the atmosphere. For their analysis, the researchers used 10 days of tropical ocean data from this simulation, which was then coarse-grained to a typical GCM resolution (around 100 km) for training and evaluation.

Key Insights and Ideas

  • Subgrid Organization is Key: The most striking insight is that simply using large-scale atmospheric variables is insufficient for accurately predicting precipitation variability and extremes. The implicit learning of subgrid-scale cloud organization is crucial.
  • "Org" as a Learned Metric: The "org" metric, learned by the autoencoder, effectively captures the complex spatial structure of convection that influences precipitation. It's not a predefined physical parameter but a data-driven representation of organization.
  • Temporal Memory of Organization: The "org" metric exhibits significant temporal correlation, meaning it can be predicted from its own past values. This highlights the "memory" effect in convective organization, where past states influence future ones, a concept previously hypothesized but now quantified.
  • Predictability of "Org": The "org" metric can be predicted from historical coarse-scale variables, suggesting that GCMs could potentially incorporate this information even without direct high-resolution data.

Key Results

  • Baseline-NN Performance: The baseline-NN, using only coarse-scale variables, could reasonably predict the average precipitation behavior but struggled significantly with its variability and extremes. Its overall prediction accuracy (R²) was around 0.45.
  • Org-NN Performance: When informed by the learned "org" metric, the Org-NN showed a dramatic improvement in predicting precipitation. It accurately captured precipitation extremes and spatial variability, achieving an R² of approximately 0.9.
  • "Org" Prediction: The "org" metric itself demonstrated good predictability from historical coarse-scale data and, more importantly, from its own past values. Using a simple autoregressive model (predicting current "org" from past "org"), R² values exceeded 0.95 for the learned organization variables.

Limitations and Future Work

  • Data Domain: The study focused on tropical ocean regions. While the findings are significant, extending this to land surfaces, which have different land-atmosphere interactions, would be valuable.
  • Interpretability of "Org": While the "org" metric is effective, fully understanding the precise physical processes it represents remains an area for future research. The authors suggest exploring the physical processes and causal mechanisms behind the temporal memory of "org."
  • Integration into GCMs: The paper demonstrates the potential of this data-driven approach. Future work will involve integrating these learned organization metrics and parameterizations into operational climate models to improve their predictive capabilities.
  • Generalizability: Investigating whether a common set of "org" variables can be learned and used effectively across different parameterizations (convection, radiation) is an important next step.

Summary updated

Abstract

Accurate prediction of precipitation intensity is crucial for both human and natural systems, especially in a warming climate more prone to extreme precipitation. Yet, climate models fail to accurately predict precipitation intensity, particularly extremes. One missing piece of information in traditional climate model parameterizations is subgrid-scale cloud structure and organization, which affects precipitation intensity and stochasticity at coarse resolution. Here, using global storm-resolving simulations and machine learning, we show that, by implicitly learning subgrid organization, we can accurately predict precipitation variability and stochasticity with a low-dimensional set of latent variables. Using a neural network to parameterize coarse-grained precipitation, we find that the overall behavior of precipitation is reasonably predictable using large-scale quantities only; however, the neural network cannot predict the variability of precipitation (<i>R</i><sup>2</sup> ∼ 0.45) and underestimates precipitation extremes. The performance is significantly improved when the network is informed by our organization metric, correctly predicting precipitation extremes and spatial variability (<i>R</i><sup>2</sup> ∼ 0.9). The organization metric is implicitly learned by training the algorithm on a high-resolution precipitable water field, encoding the degree of subgrid organization. The organization metric shows large hysteresis, emphasizing the role of memory created by subgrid-scale structures. We demonstrate that this organization metric can be predicted as a simple memory process from information available at the previous time steps. These findings stress the role of organization and memory in accurate prediction of precipitation intensity and extremes and the necessity of parameterizing subgrid-scale convective organization in climate models to better project future changes of water cycle and extremes.
BibTeX
@inproceedings{shamekh2023implicit,
  author = {Sara Shamekh and Kara D. Lamb and Y. Huang and Pierre Gentine},
  title = {{Implicit learning of convective organization explains precipitation stochasticity}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2023},
  doi = {10.1073/pnas.2216158120},
}
Figure