[Paper Review] Probabilistic Emulation of a Global Climate Model with Spherical DYffusion

Salva Rühling Cachay, Brian Henn+3 more|arXiv (Cornell University)|Jun 21, 2024

Climate Change Policy and EconomicsEconomics, Econometrics and Finance3 citations

TL;DR

This paper introduces Spherical DYffusion, a novel probabilistic generative model that emulates a global climate model with high accuracy and physical consistency over 10-year simulations at 6-hourly timesteps. By adapting the dynamics-informed diffusion framework to spherical geometry using spherical harmonics, it achieves state-of-the-art performance in climate emulation, outperforming baselines in ensemble spread and long-term statistical fidelity while maintaining computational efficiency.

ABSTRACT

Data-driven deep learning models are transforming global weather forecasting. It is an open question if this success can extend to climate modeling, where the complexity of the data and long inference rollouts pose significant challenges. Here, we present the first conditional generative model that produces accurate and physically consistent global climate ensemble simulations by emulating a coarse version of the United States' primary operational global forecast model, FV3GFS. Our model integrates the dynamics-informed diffusion framework (DYffusion) with the Spherical Fourier Neural Operator (SFNO) architecture, enabling stable 100-year simulations at 6-hourly timesteps while maintaining low computational overhead compared to single-step deterministic baselines. The model achieves near gold-standard performance for climate model emulation, outperforming existing approaches and demonstrating promising ensemble skill. This work represents a significant advance towards efficient, data-driven climate simulations that can enhance our understanding of the climate system and inform adaptation strategies.

Motivation & Objective

• To address the challenge of generating physically consistent, long-term climate simulations using data-driven models.
• To overcome the limitations of standard diffusion models in high-dimensional, spherical climate data with long inference rollouts.
• To develop a probabilistic emulator that produces reliable ensemble simulations for uncertainty quantification in climate projections.
• To enable efficient inference by capping computational overhead to less than 3× that of deterministic models.
• To demonstrate that weather forecast skill does not predict long-term climate accuracy, challenging current ML training paradigms.

Proposed method

• Adapts the DYffusion framework—originally designed for Euclidean data—to spherical geometry using spherical harmonics for global field representation.
• Employs a dynamics-informed diffusion model that incorporates physical constraints into the noise prediction process, improving stability and physical consistency.
• Uses a spherical Fourier neural operator (SFNO) as the backbone architecture to handle the spherical topology of atmospheric fields.
• Applies a conditional diffusion process where each time step is predicted based on the previous state and time-varying boundary conditions (e.g., SST, sea ice).
• Reduces sampling cost by limiting the number of forward passes to less than 3× that of a deterministic model, enabling long rollouts.
• Trains the model on a large-scale dataset of global climate simulations from the FV3GFS model, using a denoising loss function optimized for long-term statistics.

Experimental results

Research questions

• RQ1 Can a diffusion-based generative model produce stable, physically consistent 10-year climate simulations on a global scale?
• RQ2 Does optimizing for short-term weather forecast accuracy correlate with accurate long-term climate statistics in machine learning emulators?
• RQ3 How does the spherical geometry of Earth's atmosphere affect the performance of diffusion models in climate emulation?
• RQ4 Can a probabilistic model outperform deterministic surrogates in ensemble spread and uncertainty quantification for climate projections?
• RQ5 To what extent does the dynamics-informed design improve long-term statistical fidelity compared to standard diffusion models?

Key findings

• The model achieves stable 10-year simulations at 6-hourly timesteps with minimal drift in long-term climate statistics, significantly outperforming deterministic and probabilistic baselines in time-mean RMSE.
• Spherical DYffusion reduces 10-year time-mean RMSE by up to 40% compared to baseline models across multiple atmospheric fields, including surface pressure and total water path.
• The ensemble mean RMSE and CRPS of Spherical DYffusion are consistently lower than those of probabilistic baselines, with spread skill ratios closer to one, indicating better-calibrated uncertainty estimates.
• Despite strong weather forecast performance (low 5-day RMSE), DYffusion shows no correlation with long-term climate biases, confirming that weather skill is not a reliable proxy for climate accuracy.
• The model’s ensemble performance closely matches a 10-member reference ensemble for surface pressure but shows a gap for total water path, indicating field-dependent challenges in uncertainty representation.
• The method maintains computational efficiency, requiring less than 3× the forward passes of a deterministic model, enabling feasible long-horizon inference.