Title:Climate in a Bottle: Towards a Generative Foundation Model for the Kilometer-Scale Global Atmosphere

Abstract:AI emulators offer a path to compressing, boosting limited ensembles, and improving the latency of interacting with petabyte-scale climate prediction data. However, prevailing auto-regressive paradigms offer limited flexibility, and are challenging to train on climate time horizons due to drifts, instabilities and component-coupling challenges. Conditionally generative models offer an appealing alternative. In this context we demonstrate a generative diffusion-based framework -- Climate in a Bottle (cBottle) -- for emulating global km-scale climate simulations and reanalysis on the equal-area HEALPix grid. cBottle consists of two model stages: a globally-trained coarse-resolution image generator that generates 100km (50k-pixel) fields given monthly average sea surface temperatures and solar conditioning, followed by a locally-trained 16x super-resolution stage that generates 5km (12.5M-pixel) fields; global super-resolution is made affordable using an overlapping patch-based multi-diffusion. Overall, cBottle shows promise as an emulator across a battery of climate model diagnostics, including diurnal-to-seasonal scale variability, large-scale modes of variability, tropical cyclone statistics, and trends of climate change and weather extremes. Moreover, cBottle is a step towards a foundation model, by bridging multiple data modalities (reanalysis and simulation) with corresponding utility beyond emulation to tasks such as zero-shot bias correction, climate downscaling, and channel in-filling.