BEGIN:VCALENDAR VERSION:2.0 PRODID:-//pretalx//pretalx.com//juliacon-2026//speaker//FRX3ZM BEGIN:VTIMEZONE TZID:CET BEGIN:STANDARD DTSTART:20001029T040000 RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=10 TZNAME:CET TZOFFSETFROM:+0200 TZOFFSETTO:+0100 END:STANDARD BEGIN:DAYLIGHT DTSTART:20000326T030000 RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3 TZNAME:CEST TZOFFSETFROM:+0100 TZOFFSETTO:+0200 END:DAYLIGHT END:VTIMEZONE BEGIN:VEVENT UID:pretalx-juliacon-2026-GB8WXW@pretalx.com DTSTART;TZID=CET:20260812T152000 DTEND;TZID=CET:20260812T153000 DESCRIPTION:Traditionally\, climate models are difficult to run for end use rs\, and even harder to customize or interface with machine learning. We w ant to change that. Here\, we present the ongoing development of SpeedyWea ther.jl: A general circulation model that’s differentiable\, GPU-capable and ready for machine learning integration. SpeedyWeather.jl is a spectra l atmospheric general circulation model with an everything-flexible attitu de. In this talk\, we will give an overview of SpeedyWeather.jl’s develo pment of the last year\, in which we worked towards differentiability with Enzyme\, GPU-capability with KernelAbstractions and Reactant and rewrote our parametrizations for better performance and more customisability. DTSTAMP:20260502T094000Z LOCATION:Room 3 SUMMARY:SpeedyWeather.jl: Towards a differentiable and GPU-capable general circulation model - Maximilian Gelbrecht\, Milan Klöwer URL:https://pretalx.com/juliacon-2026/talk/GB8WXW/ END:VEVENT BEGIN:VEVENT UID:pretalx-juliacon-2026-UETBSG@pretalx.com DTSTART;TZID=CET:20260812T160000 DTEND;TZID=CET:20260812T161000 DESCRIPTION:Global land surface and hydrological models are crucial compone nts of Earth System Models (ESMs). In addition to providing realistic boun dary conditions for the atmosphere and ocean components\, they also play a key role in understanding Earth’s changing energy imbalance and the res ponse of the terrestrial carbon and water cycles to anthropogenic climate change. Unlike atmosphere and ocean models\, however\, land models lack a fluid dynamical core and rely heavily on empirical parameterizations to re present many key processes. As such\, there is a continued need for a new generation of land models which can facilitate the incorporation of data-d riven components. Here we present Terrarium.jl\, a Julia-based land modeli ng framework for GPU-accelerated and automatically differentiable simulati ons of soil\, snow\, and vegetation dynamics\, along with their correspond ing land-atmosphere exchange fluxes. We highlight how Julia’s key featur es enable unprecedented modularity in the model design and seamless GPU pa rallelization through KernelAbstractions.jl. We further demonstrate the va lue of GPU acceleration and differentiability through a series of performa nce benchmarks and sensitivity analyses. We also detail our initial experi ments in achieving stable coupling to a reduced-complexity atmosphere mode l\, SpeedyWeather.jl. DTSTAMP:20260502T094000Z LOCATION:Room 3 SUMMARY:Terrarium.jl: Fully differentiable and GPU-accelerated land modelin g at all scales in Julia - Brian Groenke\, Maximilian Gelbrecht URL:https://pretalx.com/juliacon-2026/talk/UETBSG/ END:VEVENT BEGIN:VEVENT UID:pretalx-juliacon-2026-U9ZWXZ@pretalx.com DTSTART;TZID=CET:20260812T172000 DTEND;TZID=CET:20260812T173000 DESCRIPTION:Hybrid climate modelling combines numerical models with machine -learned components. We present the development of multiple machine-learne d surface climate processes and their integration into the climate model S peedyWeather.jl using PyTorch and Lux.jl. Despite the offline training\, t he hybrid model is designed to generalise in space and to different climat es. We address speed vs. accuracy tradeoffs using SymbolicRegression.jl an d discuss online learning with Enzyme.jl. DTSTAMP:20260502T094000Z LOCATION:Room 3 SUMMARY:A learned surface roughness scheme for climate prediction in Speedy Weather.jl - Greg Munday\, Maximilian Gelbrecht\, Milan Klöwer\, Niklas V iebig URL:https://pretalx.com/juliacon-2026/talk/U9ZWXZ/ END:VEVENT BEGIN:VEVENT UID:pretalx-juliacon-2026-W7FRKU@pretalx.com DTSTART;TZID=CET:20260813T163000 DTEND;TZID=CET:20260813T164500 DESCRIPTION:Fortran climate models are being adapted to GPUs by automatical ly translating loop-by-loop into a kernel. In Julia\, we have more flexibi lity to develop the climate model [SpeedyWeather.jl](https://github.com/Sp eedyWeather/SpeedyWeather.jl) for the GPU. Many parts are easy to accelera te\, leverage multiple dispatch on the GPU and a high level of kernel fusi on for modularity and performance\, while being optionally hardware-specif ic. The spherical harmonic transforms remain a complex bottleneck but we e mploy a multi-algorithm approach with custom linear algebra kernels using Reactant\, Fourier and Legendre transforms. DTSTAMP:20260502T094000Z LOCATION:Room 3 SUMMARY:The GPU acceleration of SpeedyWeather.jl\, the friendly and flexibl e climate model - Milan Klöwer\, Niklas Viebig\, Maximilian Gelbrecht URL:https://pretalx.com/juliacon-2026/talk/W7FRKU/ END:VEVENT BEGIN:VEVENT UID:pretalx-juliacon-2026-EGUEJP@pretalx.com DTSTART;TZID=CET:20260814T111500 DTEND;TZID=CET:20260814T113000 DESCRIPTION:Climate models rely on parameterizations that are traditionally tuned manually. We present a differentiable calibration framework using E nzyme.jl to compute exact reverse mode gradients of energy-balance diagnos tics in SpeedyWeather.jl. By batching single- timestep gradients across ch aotic dynamics\, we enable systematic\, reproducible optimization of short wave radiation parameters\, establishing an extensible workflow for object ive calibration in Earth system models. DTSTAMP:20260502T094000Z LOCATION:Room 6 SUMMARY:Differentiable Climate Modeling: Calibrating SpeedyWeather with Enz yme - Niklas Viebig\, Milan Klöwer\, Maximilian Gelbrecht\, Greg Munday\, Brian Groenke URL:https://pretalx.com/juliacon-2026/talk/EGUEJP/ END:VEVENT END:VCALENDAR