Research and Development within the Scientific Computing Branch of NOAA’s Global Systems Laboratory

The NOAA Global Systems Laboratory, Earth Prediction Advancement Division, Scientific Computing Branch team works on approaches to facilitate development of cloud-resolving, Earth system prediction systems suitable for the next generation of exascale high performance computing (HPC), including exploration of machine learning (ML) algorithms for use within our systems or partial to full replacement of our systems for improved performance and reduced computational cost.

Our work is divided into two main categories: incremental - shorter term and innovative - longer term challenges. Work related to incremental changes focuses on existing NOAA algorithms and improvement of their performance on different architectures (e.g. adapting existing codes to run on GPUs). The more innovative aspects focus on development and evaluation of new algorithms and approaches to environmental modeling that simultaneously improve prediction accuracy, performance, and portability. For the purpose of exploring high-resolution numerics for fire weather and urban applications we have developed the GeoFLuid Object Workflow (GeoFLOW). GeoFlow is a C++ framework with convective (and other) dynamics, high order truncation, quantifiable dissipation, an option to use a variety of 2D and 3D grids, and excellent strong scaling and on-node properties. In parallel with GeoFlow development we are evaluating other available Large Eddy Simulation models, including the hybrid CPU/GPU accelerated FastEddy model. We are also actively evaluating use of the ML-based emulators for either different components of the Earth system prediction models or full replacement.

A large portion of our research and development activities involves building federated and unified workflows to facilitate both the effective use of distributed computing resources as well as easy configuration for nontrivial workflow applications in research and operations. Finally, we are evaluating viability of GT4Py Domain Specific Language (DSL) as an option to make our physics suite GPU ready.

A comprehensive summary of these research and development activities will be presented.