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

TL;DR

This work presents Climate in a Bottle (cBottle), a diffusion-based, two-stage framework that emulates kilometer-scale global atmospheric states on a 5 km HEALPix grid. By training on ERA5 reanalysis and ICON convection-resolving simulations, it achieves fast, interactive, multimodal climate state generation with capabilities for downscaling, channel infilling, and cross-dataset translation, all while enabling user-guided sampling of extremes via classifier guidance. The authors validate cBottle across diurnal and seasonal cycles, NAM, ENSO responsiveness, and tropical cyclone guidance, demonstrating its potential as a foundation model for an interactive Earth digital twin. Key contributions include diffusion on the sphere with a large-scale noise schedule, multi-diffusion for memory efficiency, and masked training to bridge diverse climate datasets, enabling zero-shot bias correction and flexible data synthesis at unprecedented resolution.

Abstract

Climate modeling is reaching unprecedented resolution, producing petabytes of data. AI climate model emulators offer a path to computationally cheap analysis, enabling new scientific insight and scenario planning. Recent advances show promise in faithfully emulating climate data. However, prevailing auto-regressive paradigms are difficult to train on climate time horizons due to drifts, instabilities, and component-coupling challenges. They are hard to scale to high resolution and require sifting through troves of output to identify rare extremes of interest. We present Climate in a Bottle (cBottle), a generative diffusion-based framework emulating global 5 km climate simulations and reanalysis on the HEALPix grid. cBottle samples directly from the full distribution of atmospheric states, avoiding auto-regressive rollout, and is the first to reach this 12.5M-pixel global resolution. It consists of two stages: a coarse-resolution generator conditioned on sea surface temperatures and solar position, followed by a patch-based 16x super-resolution stage. cBottle passes a battery of tests, including diurnal-to-seasonal variability, large-scale modes of variability, tropical cyclone statistics, and trends of climate change and weather extremes. It is a step toward a foundation model: bridging data modalities (reanalysis and simulation), enabling zero-shot bias correction, downscaling, and data infilling. It also enables new interactivity via guided diffusion. For example, we train a tropical cyclone (TC) classifier alongside the generator, guide towards TC states, and obtain physically credible samples. This opens the door to guidance methods for a wide array of user queries and new ways of interacting with climate data.

Figures (35)

• Figure 1: Climate in a Bottle: A generative foundation model for kilometer-scale climate simulation. (a) Training phase: Petabytes of high-resolution climate data from ICON global cloud resolving simulations and ERA5 reanalysis are compressed into a compact neural network representation ($\sim$ few GB). (b) Inference phase: The trained model generates realistic 5km-resolution climate states from minimal conditional inputs (time, location, sea surface temperature) through a two-stage process: Coarse global generation (100km) followed by patch-based super-resolution. (c) Applications: The foundation model enables diverse climate science applications including interactive what-if scenarios, bias correction between datasets, channel in-filling for missing variables, statistical downscaling, multimodal learning across data sources, and high-fidelity climate emulation—all achievable in minutes rather than hours and accessible without supercomputing resources.
• Figure 2: High resolution (HPX1024) samples from the ICON data (left) and synthesized by cBottle given sea surface temperature and time conditionings (right). Note these are not identically paired but use the same input SST condition for 2025-03-16T15:00:00. For each field the same colormap is used, but to highlight the visual similarity we do not include color-bars. More quantitative comparisons will follow. See Figure \ref{['fig:si:1']} for more fields.
• Figure 3: Fidelity to conditioning. Amplitude of the diurnal cycle of precipitation for (a) cBottle and (b) ERA5. DJF-JJA Arctic sea ice concentration computed over the test period for (c) cBottle and (d) ERA5. (e) cBottle's secular temperature trend during the AMIP period compared to ERA5. (f) User-provided hurricane conditioning. Minima of pressure shown in blue are linked to TCs generated by cBottle using classifier guidance. The input to the inference is the location of the cyclones, shown as turquoise boxes.
• Figure 4: Internal variability. (a,c) Tropical cyclone occurrence probability derived from ERA5 and cBottle (1980–2018), presented as heatmaps normalized per $km^2$. (b,d) The Northern Annular Mode (NAM) index for cBottle samples with overfitting (b) and large-noise-level regularization (d). The NAM indices are computed by projecting the daily Z1000 anomalies onto the NAM loading pattern (Fig \ref{['fig:si:nam']}. For convenience, the CRPS is computed by comparing the 0 UTC from the ground truth index to the 24-member ensemble of cBottle samples taken from ever hour of the same day.
• Figure 5: Multi-modal inference methods. The original ERA5 data for vertically integrated liquid water (CLLVI) (a) and $dbR=10log(precip)$ (b). (c-d) These fields translated to the ICON distribution. (e-g) The missing radiation channels infilled by cBottle-3d: outgoing shortwave (RSUT), surface downwelling shortwave (RSDS), and outgoing longwave (RLUT) (h) The downscaled dbR for this same example. This is computed by applying cBottle-SR after translation from ERA5 to ICON. These are shown for an atmospheric river example from 2006-07-11 12UTC.