Recent advances in machine learning (ML) approaches to weather forecasting have led to fully-ML forecasts at high resolution and medium-range lead times. In particular, NVIDIA has recently developed a model--Spherical Fourier Neural Operators (SFNO)--with performance comparable to state-of-the-science dynamical forecast models. Computational cost gives SFNO a significant advantage relative to traditional forecast approaches: it is approximately 10,000 times faster than traditional dynamical models, which permits the generation of large forecast ensembles. The flip side of this advantage is that SFNO ensembles can produce data at rates and quantities that can even challenge high-performance computing (HPC) centers. In-situ analysis and data reduction can potentially solve this issue.

We implement the SFNO model as an algorithm in the Toolkit for Extreme Climate Analysis (TECA) to facilitate complex in-situ analyses. TECA is a collection of climate and weather data analysis algorithms designed to make it easy for users to develop complex data analysis pipelines that take full advantage of HPC resources (i.e., thousands of nodes each with multicore processors and multiple GPUs). With SFNO as a TECA algorithm, we can compose data analysis pipelines that make use of SFNO data as it is produced--without needing an intermediate step of storing data to disk. We utilize TECA to drive ensembles of SFNO simulations, to detect weather phenomena like atmospheric rivers and tropical cyclones within those simulations, and to calculate performance metrics for comparison with observations and with other forecasting approaches. This presentation describes the implementation of SFNO within TECA and reports on the assessment of weather phenomena in SFNO. Preliminary analyses indicate that the physical representation of weather phenomena in SFNO is comparable to the representation in traditional dynamical models.