Machine Learning Emulation of Gravity Wave Drag in Numerical Weather Forecasting.

Area of Science:

• Atmospheric Science
• Computational Science
• Machine Learning

Background:

• Operational weather forecasting relies on parameterization schemes for complex physical processes.
• Nonorographic gravity wave drag is a crucial parameterization in weather models.
• Existing schemes can be computationally intensive, limiting forecast resolution and speed.

Purpose of the Study:

• To evaluate machine learning (ML) as an accelerator for weather forecasting parameterization schemes.
• Specifically, to develop and assess ML emulators for nonorographic gravity wave drag parameterization.
• To determine the accuracy and computational efficiency of ML emulators compared to traditional schemes.

Main Methods:

• Training ML emulators on an increased complexity version of the existing nonorographic gravity wave drag parameterization scheme.
• Evaluating emulator stability and accuracy across various forecasting timescales, including seasonal and medium-range.
• Comparing the performance of ML emulators against the operational parameterization scheme on both CPU and GPU hardware.

Main Results:

• ML emulators demonstrate stable and accurate results up to seasonal forecasting timescales.
• More complex neural networks generally yield more accurate emulators.
• Emulators trained on higher complexity schemes produce more accurate forecasts, outperforming the operational scheme in medium-range predictions.
• On current CPU hardware, emulators have comparable computational cost but are limited by data movement.
• On GPU hardware, ML emulators achieve a 10x speedup compared to the operational CPU-based scheme.

Conclusions:

• Machine learning offers significant potential to accelerate weather forecasting systems.
• ML emulators for nonorographic gravity wave drag can enhance forecast accuracy and efficiency, particularly on GPU architectures.
• Further development could lead to faster and more precise operational weather predictions.