Navid Constantinou: High-resolution Earth system modelling with an accelerator-agnostic, unified framework for the atmosphere and ocean
Rapid hardware evolution is reshaping scientific computing. GPUs led the transition from CPU-only systems to accelerator-based supercomputing. The landscape now includes TPUs and increasingly specialized AI-oriented architectures with programming models unlike conventional HPC. These systems offer enormous computing capacity for high-resolution Earth system models, but exploiting it requires scientific software that is performance-portable across architectures rather than rewritten for each new machine.
We describe a unified framework for atmosphere, ocean, and Earth system modelling which is differentiable, portable to CPUs, GPUs, and emerging architectures, and provides a modern, user-first scripting interface in the Julia programming language. The framework is based on Oceananigans, a system originally developed for ocean modelling, such that the atmosphere (Breeze) and the coupled modelling system (NumericalEarth) re-use Oceananigans grids, finite volume operators, algorithm and kernel design, abstractions, novel WENO-based advection schemes, turbulence closures, and user interface design. The result is a unified system that accelerates development, facilitates coupling, reduces duplicated work on numerical methods and constructions required for both atmosphere and ocean dynamical cores, facilitates collaboration across disciplines, supports hybrid ML-physics modelling and variational data assimilation, and achieves high performance on a variety of hardware including GPUs. We illustrate the potential of this framework by presenting results from atmospheric large eddy simulations of the marine boundary layer in the presence of submesoscale fronts and also results from coupled atmosphere-ocean large eddy simulations for the same setup.