Developing ECMWF’s portable next-generation atmospheric dynamical core

Numerical weather prediction directly benefits from advancements in the accuracy, efficiency, and scalability of the atmospheric model. We present the development of a performance-portable high-level Python-based framework for the next-generation ECMWF global atmospheric dynamical core, enabling simulations at unprecedented numerical resolutions. This new model framework, called the Portable Model for Multi-Scale Atmospheric Prediction (PMAP), is an advancement of the Finite-Volume Module (FVM) originally developed in Fortran at ECMWF.

A key feature of the global PMAP is its implementation with the latest version of the GridTools for Python (GT4Py) domain-specific library, named gt4py.next. This library is tailored to the efficient implementation of conservative finite-volume discretization methods that support, among others, ECMWF’s operational octahedral grid. Co-developed with various Swiss partners, the gt4py.next library itself is under continuous extension, optimization, and refinement alongside the PMAP. The model runs distributed across multiple nodes, enabling large-scale simulations on modern supercomputers with accelerators.

We present recent model validation results and provide an analysis of its performance, portability, and scalability on latest European supercomputing platforms.