BEGIN:VCALENDAR PRODID:-//eluceo/ical//2.0/EN VERSION:2.0 CALSCALE:GREGORIAN BEGIN:VEVENT UID:3a1b7e8f7d2591c0c260bc7ccc46e493 DTSTAMP:20261027T222403Z SUMMARY:CCMI Seminar – Boris Bonev (NVIDIA Zurich\, École Polytechnique Paris) DESCRIPTION:Speaker: Boris Bonev (NVIDIA Zurich\, École Polytechnique Par is)\nTitle: Geometric Neural Operators for Machine Learning Probabilistic Weather Prediction \nAbstract: Machine learning has recently enabled fas t and accurate surrogate models for numerical weather and climate predicti on. However\, many current architectures struggle on extreme events\, intr oduce spurious artifacts or instabilities when they ignore the spherical g eometry of the Earth\, and are often tightly coupled to a particular input grid or resolution\, limiting flexibility for downstream applications. Th ese shortcomings highlight a mismatch between black‑box models and more traditional numerical methods\, which encode underlying physical and geome tric properties. Neural operators address part of this gap by learning sol ution operators of partial differential equations in a grid‑agnostic man ner\, yet standard Fourier Neural Operators (FNOs) break down on the spher e\, where the flat‑geometry Fourier transform induces artifacts and exce ssive dissipation.\nThis talk presents geometric neural operators that res pect the symmetries and topology of the sphere\, in the spirit of the geom etric deep learning blueprint\,  with a particular focus on applications to the chaotic dynamics of Earth’s atmosphere. The Spherical Fourier Neu ral Operator (SFNO) is formulated via Driscoll–Healy’s convolution the orem\, linking spherical harmonic transforms to group convolutions on the sphere and yielding rotationally equivariant\, grid-invariant models capab le of stable\, year‑long autoregressive rollouts with physically plausib le dynamics. Building on insights from hyperbolic PDE solvers\, a principl ed framework is introduced for incorporating localized integral and differ ential operators into neural operator architectures\, using provably conve rgent differential layers and quadrature‑based discrete–continuous con volutions on both Euclidean and spherical geometries to capture sharp fron ts and local extremes. To complement convolutional approaches\, a generali zed attention mechanism for spherical domains is developed\, allowing Tran sformer‑style architectures to natively process data on the two‑dimens ional sphere. This discretization‑agnostic spherical attention incorpora tes numerical quadrature weights into a continuous spherical formulation\, producing geometrically faithful attention that is approximately rotation ally equivariant.\n\nThe practical impact of these ideas is demonstrated w ith FourCastNet 3\, a scalable probabilistic weather forecasting system ca st as a hidden Markov model built from spherical signal‑processing primi tives. FourCastNet 3 is trained on 1000+ GPUs using flexible domain-parall elism paradigms inspired from traditional HPC methods. The resulting metho d matches or exceeds leading conventional ensemble systems and state‑of ‑the‑art diffusion models while delivering 8–60× faster forecasts\, positioning it as a robust foundation for next‑generation weather and c limate modeling and downscaling systems.\n\nBio: Boris is an incoming Prof essor at École Polytechnique (joining November 2026) and Senior Research Scientist at NVIDIA\, working at the frontier of AI4Science by uniting rig orous numerical analysis\, high‑performance GPU computing\, and large‑ scale machine learning. He pioneered Spherical Neural Operators and the Fo urCastNet family of AI weather models\, which enable 7\,000‑member ensem bles and are deployed by agencies such as NOAA\, ECMWF\, and MetDesk for n ext‑generation global forecasting. He leads research on operator learnin g for hyperbolic PDEs and data‑driven digital twins of the Earth system\ , developing principled neural operators that scale to 1\,000+ GPU cluster s while preserving stability and physical structure. He completed his PhD in applied mathematics at EPFL. URL:https://www.imperial.ac.uk/events/215762/ccmi-seminar-boris-bonev-nvidi a-zurich-ecole-polytechnique-paris/ DTSTART;TZID=Europe/London:20261103T140000 DTEND;TZID=Europe/London:20261103T150000 LOCATION:Function space\, UCL AI Centre\, 90 High Holborn\, London\, WC1V 6 BH\, United Kingdom END:VEVENT BEGIN:VTIMEZONE TZID:Europe/London BEGIN:STANDARD DTSTART:20261103T140000 TZNAME:GMT TZOFFSETTO:+0000 TZOFFSETFROM:+0000 END:STANDARD END:VTIMEZONE END:VCALENDAR