Taimoor Sohail: Disentangling intrinsically variability from long-term climate change using GPU-powered large ensembles
High-resolution ocean models that resolve eddies are computationally expensive, which historically limits most simulations to a single long run rather than an ensemble of many. We address this using Oceananigans.jl and NumericalEarth.jl, a GPU-based ocean modeling framework that runs up to 50x faster than traditional CPU-based models. This speedup makes it possible to generate large ensembles of eddy-resolving simulations, something that is otherwise impractical.
We run a 50 year global ocean simulation at 1/6° resolution using repeat-year forcing (RYF). From this base, we branch off 50 separate simulations, each starting from slightly different initial conditions — an approach known as a single-model initial-condition large ensemble (SMILE). By comparing key metrics such as ocean heat content, ocean and sea-ice circulation, and sea surface temperature across all 50 runs, we measure how much variation arises from natural, internal variability alone, and how well this can be distinguished from the signal of long-term climate change.
This talk covers three parts: the practical challenges and advantages of GPU-based ocean modeling with Oceananigans.jl and NumericalEarth.jl; the design of our experimental setup and ensemble workflow; and the key findings from our large-ensemble analysis, focused on how clearly a climate change signal can be separated from natural variability at this resolution.