Latent Spaces for Weather & Climate Prediction | Pierre Gentine | AI + Environment Summit 2026

Keynote "Lost (and Found) in Latent Land: How latent spaces can improve our prediction and understanding of weather and climate" by Prof. Dr. Pierre Gentine (Columbia University, co-founder of Celest.Science) at the AI + Environment Summit 2026. Recorded 30 September 2026 at Innovationspark Zürich, Dübendorf, Switzerland. Q&A included. AI weather models usually learn in the space of physical variables. This keynote argues that a learned latent space can do much more than compress the data. It walks through one argument in five steps: • Latent space as compression: finding the low-dimensional structure of weather and climate data • Latent space as dynamics: a latent Koopman model for the predictable, coherent part of the flow plus a conditional diffusion model for the stochastic fluctuations, giving sub-seasonal forecasts 44 days ahead on par with NCEP and ECMWF at a fraction of the cost • Latent space as a physical representation: training with a latent-space loss keeps long rollouts sharp instead of blurry and better preserves multivariate relationships, checked with geostrophic balance at 500 hPa against ERA5 • Latent space as an interface: embedding model states and heterogeneous observations (satellites, ground stations, remote sensing) into one unified latent space • Latent space as an inference space: data assimilation as conditional generation, with the SLAMS framework (diffusion-based latent data assimilation) and latent 3D-Var / 4D-Var Latent-space assimilation turns out cheaper and more accurate than model-space assimilation, because the latent space is quasi-linear and quasi-Gaussian. The talk closes with where it still fails (patchy convective rain) and with the limitation that latent data assimilation can only recover structures the latent space can represent. Recording note: Zoom recording of the slides with the speaker's video where available. The introduction is not included; the Q&A is. The AI + Environment Summit is a full-day summit to inspire, ignite and innovate work in AI for the environment, organised by PhD fellows and postdoctoral researchers from ETH AI Center, EPFL, UZH and other institutions. It is run by a registered Swiss non-profit. Website: https://ai-environment-summit.com/ LinkedIn:   / ai-environment-summit-zurich   Bluesky: https://bsky.app/profile/ai-env-summi... Contact: contact@aiesummit.com Chapters: 00:00 The AI revolution in weather and climate modelling 04:00 Prediction has improved, understanding has not: why latent spaces 06:08 Turbulence, coherent structures and why they are low-dimensional 08:32 Latent Koopman model plus conditional diffusion for the stochastic part 11:46 Sub-seasonal forecasts 44 days ahead, on par with NCEP and ECMWF 14:55 Latent diffusion as a foundation model for physics (Polymathic, The Well) 17:48 Training in latent space vs model space: sharp vs blurry rollouts 20:49 Physical consistency check: geostrophic balance against ERA5 22:41 Data assimilation as conditional generation in a unified latent space 28:27 Latent 3D-Var and 4D-Var 32:22 Background-error covariance is nearly diagonal in latent space 34:12 Where it still fails: convective rain 34:44 Conclusion 35:40 Q&A Thumbnail artwork: Alan Wharton / Better Images of AI (https://betterimagesofai.org), CC BY 4.0. #AIWeatherForecasting #ClimateModeling #DataAssimilation #DiffusionModels #MachineLearning #AIForEnvironment #AIEnvironmentSummit