Posted by Tapio Schneider, Visiting Researcher, and Yi-fan Chen, Engineering Lead, Google Research
Today, climate models are able to capture general global warming trends. However, due to uncertainties in small-scale yet globally significant processes like clouds and ocean turbulence, these models lack detailed accuracy in predicting future climate changes. For instance, predictions of when the Earth’s global mean surface temperature will increase by 2℃ relative to preindustrial times vary by 40-50 years among current models. This lack of accuracy hinders our ability to plan resilient infrastructure, adapt supply chains to climate disruptions, and assess climate-related risks for vulnerable communities. One of the main reasons for this is the dominant role of clouds in errors and uncertainties in climate predictions for the coming decades [1, 2, 3]. Clouds play a crucial role in regulating Earth’s energy balance and influencing the climate system’s response to changes in greenhouse gas concentrations through sunlight reflection and greenhouse effects. However, their small scale prevents direct resolution in current climate models. These models can only resolve motions at scales of tens to hundreds of kilometers, with some pushing towards the kilometer-scale. However, the turbulent air motions that sustain low clouds over large areas of tropical oceans occur at scales of meters to tens of meters. This significant difference in scale leads to the use of empirical parameterizations of clouds in climate models, resulting in large errors and uncertainties. Although global climate models cannot directly resolve clouds, their turbulent dynamics can be simulated in limited areas using high-resolution large eddy simulations (LES). However, the high computational cost of simulating clouds with LES has hindered widespread and systematic numerical experimentation, as well as the generation of large datasets for training parameterization schemes in coarser-resolution global climate models. In our study “Accelerating Large-Eddy Simulations of Clouds with Tensor Processing Units” published in the Journal of Advances in Modeling Earth Systems (JAMES), in collaboration with a Climate Modeling Alliance (CliMA) lead who is a visiting researcher at Google, we demonstrate the effective use of Tensor Processing Units (TPUs) for performing LES of clouds. TPUs are application-specific integrated circuits originally developed for machine learning (ML) applications. We show that with tailored software implementations, TPUs can simulate computationally challenging marine stratocumulus clouds observed in the Dynamics and Chemistry of Marine Stratocumulus (DYCOMS) field study. This successful TPU-based LES code demonstrates the utility of TPUs, with their large computational resources and tight interconnects, for cloud simulations. Over the past 20 years, climate model accuracy for critical metrics such as precipitation and energy balance at the top of the atmosphere has improved by approximately 10% per decade. Our aim is to reduce climate model errors by 50% through improved cloud representation. In this research, we specifically focus on stratocumulus clouds, which cover approximately 20% of tropical oceans and are the most prevalent cloud type on Earth. Current climate models struggle to accurately reproduce stratocumulus cloud behavior, leading to significant errors. Our work provides a more accurate representation of stratocumulus clouds for large-scale climate models. Our simulations of clouds on TPUs achieve unprecedented computational throughput and scaling, allowing for the simulation of stratocumulus clouds at speeds 10 times faster than real-time across areas close to the size of typical global climate model grid boxes. These results open up new possibilities for computational experiments and the expansion of LES datasets for training cloud parameterizations in global climate models. The LES code is written in TensorFlow, an open-source software platform developed by Google for ML applications. The code takes advantage of TensorFlow’s graph computation and Accelerated Linear Algebra (XLA) optimizations, which fully utilize TPU hardware, including high-speed inter-chip interconnects that contribute to the exceptional performance achieved. Additionally, the TensorFlow code allows for the incorporation of ML components directly within the physics-based fluid solver. We validated the code by simulating canonical test cases for atmospheric flow solvers and successfully reproducing the cloud fields and turbulence characteristics observed during the DYCOMS field campaign. This achievement is particularly challenging for LES due to rapid changes in temperature and other thermodynamic properties at the top of stratocumulus decks. With this foundation established, our next goal is to expand existing databases of high-resolution cloud simulations. These simulations will provide researchers developing climate models with better cloud parameterizations, whether they are based on physics, ML, or a combination of the two. This expansion will involve incorporating additional physical processes, such as radiative transfer, into the code. Our ultimate objective is to generate data for various cloud types, including thunderstorm clouds. This work demonstrates how ML hardware advancements can be effectively applied in other research areas like climate modeling. Our simulations provide detailed training data for critical processes like in-cloud turbulence, which are essential for accurate climate modeling and prediction. We would like to express our gratitude to the co-authors of the paper: Sheide Chammas, Qing Wang, Matthias Ihme, and John Anderson. We would also like to thank Carla Bromberg, Rob Carver, Fei Sha, and Tyler Russell for their valuable insights and contributions to this work.
Source link
