Minisymposium

Data-driven probabilistic forecasts have become a tangible possibility within just a couple of years, thanks to breakthroughs mostly driven by the tech industry and building on existing open datasets from the weather and climate community. National weather services play a crucial role in providing accurate and timely weather forecasts, essential for public safety and economic planning. The integration of machine learning (ML) presents transformative opportunities and challenges in enhancing predictive capabilities and providing novel products. We will showcase a few applications of ML which are already in operations, discuss some of the scientific and computing challenges which we have encountered, and present some early results from our efforts to build a regional data-driven forecasting model.

Recent developments in machine learning for weather forecasting have led to data-driven models that are comparable in skill to leading physics-based NWP systems. Over the past year, ECMWF has been developing its own data-driven forecasting system, the AIFS. By leveraging both data and model parallelism, AIFS can be trained across O(100) GPUs; the latest version runs at a resolution of ca. 0.25-degrees. We give an overview of AIFS and ai-models, the pipeline that has been developed by ECMWF to produce data-driven weather forecasts, and runs daily – with open data delivery - on ECMWF's HPC. In addition, we showcase early results of ongoing research efforts at ECMWF, including data-driven probabilistic ensemble forecasting, and direct observation prediction - a task that aims to produce a weather forecast solely from observational data.

The AI2 Climate Emulator (ACE) marks a significant leap in climate modeling, employing a deep learning framework to replicate the comprehensive dynamics of the FV3GFS atmospheric model efficiently. ACE incorporates a Spherical Fourier Neural Operator (SFNO) with approximately 200M parameters. Using the previous weather state and externally prescribed forcings, this model forecasts the atmospheric state 6 hours ahead, alongside diagnostics such as surface precipitation rate, and turbulent and radiative fluxes. This variable set facilitates a robust assessment of the moisture and dry air mass budgets, and allows us to incorporate constraints to conserve dry air mass and ensure a closed moisture budget. Trained on a dataset with 100 years of atmospheric states simulated by a physics-based global atmosphere model, coarsened to a resolution of 1° and eight vertical levels, ACE demonstrates the ability to conduct stable multi-decadal simulations that maintain accurate weather dynamics and seasonal cycles, closely mirroring the reference model's precipitation and temperature patterns. Notably, ACE uses 60 times less energy than the FV3GFS model, leveraging modern GPU technology for efficient inference. This study underscores the potential of machine learning in climate prediction, offering a path towards fast and accessible climate models.

In this panel discussion, the speakers of the session will be asked about their view on how machine learning for Earth system modelling will evolve in the coming years, and how machine learning and conventional models will interact, merge or coexist. The session will start with a couple of questions by the chair but will also give the audience the opportunity to ask questions and to make comments.