A Python Dynamical Core for Numerical Weather Prediction

MeteoSwiss uses the ICON model to produce high-resolution weather forecasts at kilometre scale, with GPU support enabled through an OpenACC-based Fortran implementation. While effective, this approach limits portability, maintainability, and development flexibility. Within the EXCLAIM project, we focus on the dynamical core of the model—responsible for approximately 55% of the total runtime—and explore alternatives based on a domain-specific Python framework. In particular, we reimplemented the computational stencils using GT4Py and integrated them into the existing Fortran codebase, enabling the partial replacement of key components. This hybrid approach aims to improve developer productivity and code adaptability while preserving performance. In this contribution, we present our strategy for developing software for a weather and climate model involving multiple institutions and stakeholders. We present several optimisation techniques and compare the performance of the new implementation with the original OpenACC version. Our results show improved computational efficiency alongside a substantial improvement in the development workflow. Finally, we discuss the practical challenges of integrating Python components into operational numerical weather prediction systems.