Neural general circulation models for modeling precipitation

Area of Science:

• Climate Science
• Machine Learning Applications in Meteorology

Background:

• General Circulation Models (GCMs) exhibit limitations in accurately simulating precipitation, especially extremes and diurnal patterns, impacting climate studies and human activities.
• Existing hybrid models combining physics and machine learning have not yet surpassed traditional GCMs in performance.

Purpose of the Study:

• To develop and evaluate a novel hybrid climate model using the differentiable NeuralGCM framework for improved precipitation simulation.
• To leverage direct training on satellite-based precipitation observations to enhance model accuracy.

Main Methods:

• Development of a hybrid model within the differentiable NeuralGCM framework.
• Direct training of the model using satellite-based precipitation observations.
• Evaluation of the model's performance against GCMs, reanalysis data (ERA5), and a global cloud-resolving model for climate simulation, and against the ECMWF ensemble for forecasting.

Main Results:

• The hybrid model demonstrates substantial improvements in precipitation simulation compared to existing GCMs, ERA5 reanalysis, and a global cloud-resolving model.
• The model outperforms the ECMWF ensemble in mid-range precipitation forecasting.
• The approach showcases the efficacy of training climate models directly on observational data.

Conclusions:

• The developed hybrid model represents a significant advancement in simulating precipitation, offering more reliable climate projections.
• Directly training climate models on observational data is a viable and effective strategy for enhancing their predictive capabilities.