TL;DR
PACER is an ultra-lightweight (2.1M parameters) climate emulator that pairs a deterministic 2D advection partial differential equation (PDE) backbone with a stochastic Neural SDE corrector. It generates stable, physically grounded, 10-year climate projections across 20 numerical climate models under variable greenhouse gas (GHG) forcings while providing well-calibrated uncertainty estimates.
What Problem Does This Paper Address?
Traditional general circulation models (GCMs) solve massive systems of physical partial differential equations to project multi-decadal climate scenarios. However, running these simulations is computationally grueling, often requiring supercomputers to churn for months.
While recent deep learning models excel at medium-range weather forecasting, they frequently suffer from numerical instabilities or climate drift when rolled out autoregressively over decadal timescales. Contemporary climate emulators (such as ACE or Spherical Dyffusion) mitigate some of these problems but are parameter-heavy (~200M parameters), rarely condition on time-varying greenhouse gas emissions directly, and often rely on deterministic objectives that cannot capture aleatoric climate variability.
Methods
PACER divides the climate emulation step into a two-stage hybrid framework:
[State u(t) & GHG Forcings F]
│
▼
┌──────────────────────────────────────┐
│ Stage 1: Physics Backbone (ODE) │
│ - Infer flow velocity v(t) │
│ - Solve 2D advection-forcing PDE │
│ using finite differences + dopri5 │
└──────────────────┬───────────────────┘
│ x_next (ODE state)
▼
┌──────────────────────────────────────┐
│ Stage 2: Stochastic Refinement (SDE)│
│ - Encode to latent state z_0 │
│ - Latent SDE (U-Net drift + diff.) │
│ - Sample K Brownian paths (SRK) │
│ - Decode to refined state & variance│
└──────────────────────────────────────┘
-
Deterministic Physics Backbone (ODE): Rather than letting neural nets learn transport from scratch, PACER models large-scale transport explicitly via a 2D advection-forcing PDE: $$\frac{\partial u}{\partial t} + v_x \frac{\partial u}{\partial x} + v_y \frac{\partial u}{\partial y} = S(u, F; x, y, t)$$ Flow velocities ($v_x, v_y$) are initialized from historical trends via cubic splines and finite differences, then dynamically refined using a small 2-layer $1\times 1$ convolutional MLP. The semi-discrete spatial system is solved using the Dormand–Prince (
dopri5) ODE integrator. -
Stochastic Residual Corrector (Neural SDE): To account for sub-grid process noise, unresolved physics, and systematic biases, the ODE output is encoded into a latent space and advanced using an Itô Stochastic Differential Equation: $$dz_t = f_\theta(z_t, t)\,dt + g_\theta(z_t, t)\,dW_t$$ The drift $f_\theta$ is a multiscale U-Net, while the diffusion $g_\theta$ scales stochastic Brownian increments state-dependently to capture heteroscedastic uncertainty.
-
Likelihood-Based Training: Unlike prior generative emulators trained via MSE or diffusion loss, PACER draws $K > 1$ Brownian paths and minimizes a latitude-weighted Gaussian Negative Log-Likelihood (NLL) with a half-Gaussian variance prior, curbing gradient variance ($\propto 1/K$) and preventing variance collapse.
Datasets
The authors evaluated PACER across 20 distinct physics-based climate models:
- ClimateSet (19 CMIP6 models): Spans 86-year projections (2015–2100) at monthly resolution and $250\text{ km}$ ($96 \times 144$) spatial resolution across Shared Socioeconomic Pathways (SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5).
- FV3GFS: A 10-year, 11-member ensemble from the US National Weather Service/NCEP regridded to a $1^\circ$ Gaussian grid ($180 \times 360$).
Key Insights and Ideas
- Injecting simple physics beats brute force: Atmospheric dynamics are governed by complex Navier–Stokes equations, but enforcing just basic 2D advection provides a surprisingly strong inductive bias. This allows a 2.1M-parameter model to outperform much larger purely data-driven alternatives.
- Residual SDEs fix physics mismatch: Instead of relying entirely on an imperfect idealized PDE, the Neural SDE acts strictly as a residual corrector over the ODE trajectory, learning fast sub-grid scales and noise without disrupting physical consistency.
- Cross-model zero-shot generalization: Because the core transport mechanics are explicitly modeled via advection, PACER generalizes well to unseen climate models without fine-tuning.
Key Results
- 10-Year Roll-Outs (CMIP6): Evaluated on surface air temperature (TAS), PACER consistently outperformed UNet, ConvLSTM, and Spherical Fourier Neural Operators (SFNO). On average, PACER achieved RMSE of ~0.35–0.42 K across all 19 ClimateSet models (compared to ~0.38–0.50 K for SFNO and 0.67–1.43 K for UNet).
- Multi-Level Vertical Fields (FV3GFS): On 7 vertical atmospheric levels, PACER achieved the lowest RMSE and CRPS across all vertical levels (e.g., $T_5$ RMSE of 0.349 vs. 0.379 for SFNO and 0.899 for UNet).
- Zero-Shot Transfer: When pre-trained on one climate model and tested on another (e.g., trained on GFDL-ESM4, tested zero-shot on AWI-CM-1-1-MR), PACER maintained strong roll-out stability (RMSE 0.198 vs. 0.325 for SFNO).
- Ablation Evidence: Stripping the advection ODE backbone and training the Neural SDE purely on data led to notable performance degradation across all evaluated models.
Limitations and Future Work
- Oversimplified Physics: PACER currently only enforces 2D advection. Real climate dynamics involve rotating primitive equations, moist convection, thermodynamics, radiative transfer, and ocean-atmosphere coupling.
- Coarse Resolution: Experiments were conducted on a relatively coarse $250\text{ km}$ grid, which obscures localized extreme weather phenomena (such as tropical cyclones or convective storms).
- Future Directions: Expanding the physics backbone beyond advection toward primitive equation constraints, scaling to higher resolutions, and conditioning on a broader suite of diagnostic variables (e.g., precipitation, pressure).
Summary updated