Advances in Climate Modeling with Large-Eddy Simulations
Climate models today capture global warming trends, but have uncertainties about smaller processes like clouds, leading to inaccurate predictions of future climate changes. Clouds play a big role in climate predictions since they reflect sunlight, regulate Earth’s energy balance, and affect the climate’s response to changes in greenhouse gases.
Clouds can’t be directly resolved in global climate models, so high-resolution large eddy simulations (LES) can be used to simulate their turbulent dynamics. However, the high computational cost of these simulations has slowed down progress. The new study “Accelerating Large-Eddy Simulations of Clouds with Tensor Processing Units” shows that Tensor Processing Units (TPUs) — originally designed for machine learning (ML) — can effectively perform LES of clouds.
This TPU-based LES code improves climate model accuracy and enables researchers to get more accurate cloud representations. The simulations exhibit unprecedented computational throughput and scaling, making it possible to simulate clouds with a 10× speedup over real-time evolution. The LES code is written in TensorFlow, an open-source software platform developed by Google for ML applications.
The next goal is to generate data across a variety of cloud types, such as thunderstorm clouds, to further improve climate modeling. By repurposing ML hardware for climate modeling, these simulations provide detailed training data for crucial processes like in-cloud turbulence, which are essential for climate modeling and prediction.
This research marks a significant advancement in climate modeling by enhancing predictions of climate changes, and it shows how advances in ML hardware can be repurposed effectively for other areas of research.