Probabilistic Emulation of a Global Climate Model with Spherical DYffusion

TL;DR

This work presents 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.

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.

Figures (16)

• Figure 1: Weather performance (x-axis) is not a strong indicator of climate performance (y-axis). Each dot corresponds to a distinct sample or checkpoint epoch.
• Figure 2: RMSE of 10-year time-means for a subset of important fields. The leftmost bar in the first two subplots shows the reference noise floor, determined by comparing ten independent 10-year reference FV3GFS simulations with the validation simulation. The scores computed using the mean over these ten simulations (a proxy for an "ensemble prediction") are shown in light shade. The subsequent bars show the corresponding scores for our method and the deep-learning baselines, using a 25-member ensemble for the probabilistic methods (all except ACE, which only reports scores for its single deterministic prediction). Scores computed using the ensemble-mean prediction are shown in light shade. The dark shaded bar on top indicates the performance drop when using a single member's prediction only, with error bars representing the standard deviation over the 25 different member choices. The rightmost subplot displays the average time-mean RMSE of the ML-based emulators relative to the reference across all 34 variables. On average, our method's time-mean RMSEs are $50\%$ higher than the noise floor, which is less than half the average RMSE of the next best method, ACE. When using the 25-member ensemble mean prediction, this reduces to $29.28\%$.
• Figure 3: The diagram shows how our proposed approach functions at inference time. Given an initial condition $\boldsymbol{x}_{t}$ and forcings $\boldsymbol{f}_{t:t+h}$, our method uses the DYffusion framework, integrated with two SFNO backbone networks, to generate predictions for the next $h$ time steps based on an alternation of direct multi-step forecasts and temporal interpolations. To simplify the visualization, we exclude the facts that the interpolator network, $\text{SFNO}_\phi$, is conditioned on $\boldsymbol{x}_{t}$ and $\boldsymbol{f}_{t}$ in addition to an estimate of $\boldsymbol{x}_{t+h}$. We also exclude the time-conditioning of both networks. To forecast more time steps beyond $t+h$, our method is applied autoregressively.
• Figure 4: Diagram of one of the blocks of the modified SFNO architecture for our proposed method. The full architecture consists of a sequence of 8 such blocks. Our newly introduced time-conditioning modules correspond to the Time Embedding, followed by the MLP on the right, and the scale-shift operation. Our method relies on dropout, which is part of the two-layer MLP on the top. SFNO-based baselines use the same architecture and hyperparameters without the time embedding module.
• Figure 5: Global maps of the 10-year time-mean biases of a single sample from the reference noise floor simulation, our model, and the ACE baseline for the total water path field. Each subplot reports the global mean RMSE and bias of the respective bias map. Our model reproduces biases of similar location and magnitude to the reference noise floor, suggesting they are mainly due to internal climate variability rather than model bias, while the baseline exhibits larger climate biases.