A high-resolution boost for global climate modeling

A computationally efficient statistics-based approach has made it possible to emulate global climate simulations at ultra-high spatial resolution for the first time, shows research by a KAUST-led team^[1].

“Climate simulations generated by Earth system models are indispensable for advancing understanding of climate processes, predicting future changes and developing strategies to address the challenges posed by climate change,” says KAUST postdoc Yan Song. “However, generating these simulations requires extensive computation, often taking weeks or months.”

Song, along with colleague Marc Genton and collaborator Zubair Khalid from Lahore University of Management Sciences in Pakistan, took a new look at the intricate Earth system models (ESMs) that describe global climate dynamics with a view to applying statistical methods to improve their efficiency.

“ESMs enable comprehensive and detailed climate simulations, but they are computationally expensive and require massive amounts of data storage, limiting their practicality for ultra-high-resolution applications,” says Song. “Leveraging statistical techniques, we constructed a practical complement for ESMs called a statistical emulator that captures intrinsic spatiotemporal structures of the ESMs and generates fast stochastic approximations.”

Generating simulations with ESMs involves iteratively solving a series of equations for each grid cell globally over time. A statistical emulator instead uses a much less complex statistical approach with trained parameters to generate a stochastic imitation of the simulation output. Then, just the emulator’s stochastic parameters require storage, instead of the full outputs of many climate simulations.

Song, Genton and Khalid’s emulator is based on a mathematical spherical harmonic transformation that converts spatial information on a sphere into the frequency domain to enable more efficient statistical analysis.

“Spherical harmonic transformations are useful for identifying dominant spatial variations and patterns, and enables analysis in the frequency domain, which significantly accelerates computations,” says Song.

Using their emulator approach, the researchers successfully produced emulations of simulated global surface temperatures from the newly published CESM2-LENS2 dataset at a daily timescale and a spatial resolution of 110 km — the first time emulations at such temporal and spatial resolution have been attempted.

Leveraging exascale computing resources, the researchers then extended their method to enable the stochastic emulation of global surface temperatures at an hourly timescale and spatial resolution of just 3 km, for which they have been recognized with a nomination for the Gordon Bell Prize — a prestigious award for an outstanding accomplishment in high-performance computing.

“The problem addressed in this work is of significant value to climate scientists,” Song says.