AI for climate modeling from present to future

  • * Register (or log in) to the Neural Network to add this session to your agenda or watch the replay

  • Date
    23 September 2025
    Timeframe
    17:00 - 18:00 CEST
    Duration
    60 minutes

      AI-driven weather forecast models are now more accurate and faster than the best physics-based systems. Extending these advances to seamless weather–climate prediction poses broader challenges, but progress is rapid. Several models trained on ERA5 capture historical variability and trends, with some coupled to simple AI ocean components for seasonal forecasts. The key question is whether such systems can generalize to project future climate reliably. Purely data-driven extrapolation remains elusive, but emulators of physics-based models trained across multiple climates are emerging as a promising route. These can generate ensembles of ocean-coupled simulations that are statistically consistent with their reference models but at orders-of-magnitude lower cost.

      I present results from the open-source Ai2 Climate Emulator (ACE), built on NVIDIA’s SFNO architecture. ACE emulates daily weather variability and climate at 100 km resolution, running ~1600 years/day on a single GPU—about 100× faster than comparable physics-based models. It can be trained on ERA5 or AMIP-style forcings, paired with AI downscaling to km-scale weather, or coupled to slab-ocean models to capture climate change responses. Most recently, when coupled to Samudra, a full-depth ocean emulator, ACE reproduced stable coupled climate states and realistic El Niño–Southern Oscillation variability. I conclude with remaining generalization challenges toward CMIP-type applications.

      Are you sure you want to remove this speaker?