Implicit learning of convective organization explains precipitation stochasticity

Area of Science:

• Climate science
• Atmospheric physics
• Machine learning applications in meteorology

Background:

• Accurate precipitation intensity prediction is vital for human and natural systems, particularly in a warming climate with increased extreme precipitation events.
• Current climate models struggle to accurately predict precipitation intensity and extremes due to limitations in parameterizing subgrid-scale cloud structure and organization.

Purpose of the Study:

• To investigate the impact of subgrid-scale cloud organization on precipitation intensity prediction using storm-resolving simulations and machine learning.
• To develop and evaluate a machine learning approach that implicitly learns subgrid organization for improved precipitation variability and extreme event forecasting.

Main Methods:

• Utilized global storm-resolving simulations and machine learning, specifically neural networks, to parameterize coarse-grained precipitation.
• Developed an 'organization metric' by training a machine learning algorithm on high-resolution precipitable water fields to implicitly capture subgrid organization.
• Assessed the predictive performance of the neural network with and without the organization metric, and explored its temporal dynamics using previous time steps.

Main Results:

• A neural network using only large-scale quantities could predict overall precipitation behavior but failed to capture variability (R² ≈ 0.45) and underestimated extremes.
• Incorporating the learned organization metric significantly improved predictions of precipitation extremes and spatial variability (R² ≈ 0.9).
• The organization metric exhibited significant hysteresis, indicating the influence of memory from subgrid-scale structures, and was predictable from past information.

Conclusions:

• Subgrid-scale cloud organization and memory are critical factors for accurate precipitation intensity and extreme event prediction.
• Parameterizing subgrid-scale convective organization in climate models is necessary for more reliable projections of future water cycle changes and extreme events.
• Machine learning, by implicitly learning organization, offers a promising avenue for enhancing precipitation forecasting in climate models.