Neural Emulators for Scaling Weather and Climate Models

194 concepts181 papers1 proposal

Traditional weather forecasting and climate projection rely on Numerical Weather Prediction (NWP) models and General Circulation Models (GCMs). These systems discretize the Earth's atmosphere and oceans into a three-dimensional grid, using supercomputers to iteratively solve the partial differential equations (PDEs) governing fluid dynamics and thermodynamics.

While physically rigorous, this classical approach faces a severe computational bottleneck. To capture localized, high-impact phenomena like cloud formation, convective storms, or urban microclimates, models must operate at kilometer-scale or sub-kilometer resolutions. However, increasing spatial resolution incurs an exponential computational penalty: a tenfold increase in resolution can require up to a thousand-fold increase in computing power. This constraint makes multi-decadal climate projections or large-scale probabilistic ensemble forecasting prohibitively expensive.

To bypass this bottleneck, a rapidly growing body of research focuses on neural emulation—replacing or augmenting resource-intensive numerical solvers with deep learning architectures. By learning the underlying mappings of atmospheric and oceanic states from historical reanalysis data or high-resolution simulation outputs, these models achieve speedups of several orders of magnitude.

Accelerating Sub-Grid Physics via Neural Parameterization

A primary source of computational overhead in traditional climate models is the calculation of "parameterizations"—semi-empirical subroutines designed to approximate physical processes that occur at scales smaller than a single grid cell, such as cloud microphysics, radiation transfer, and planetary boundary layer dynamics. These subroutines can consume up to 70% of a model's total processing time.

Instead of calculating these complex physical interactions from scratch at every time step, researchers use deep neural networks to emulate them. For instance, Wang et al. developed domain-aware deep neural networks to emulate planetary boundary layer parameterizations, achieving speeds up to 80 times faster than traditional formulations by designing network architectures that inherently respect the vertical structure of the atmosphere. Similarly, Novitasari and Quaas introduced an active learning framework to predict autoconversion rates—a critical sub-grid process in precipitation formation—drastically reducing the data overhead required to simulate cloud-to-rain transitions.

Why does this work? Neural networks function as highly efficient universal function approximators. Once trained on the inputs and outputs of an expensive physical subroutine, the network can predict the physical tendencies (e.g., changes in temperature or moisture) via a series of highly parallelized matrix multiplications. This shifts the computational burden from runtime PDE solving to a one-time offline training phase.

Global Data-Driven Forecasting and Spherical Geometry

Beyond emulating isolated subroutines, recent breakthroughs demonstrate that deep learning can replace the entire numerical core of global weather models. Systems like Pangu-Weather, introduced by Bi et al., utilize 3D neural networks trained on decades of historical weather data to outperform state-of-the-art operational NWP systems in deterministic forecasting accuracy, operating at a fraction of the computational cost.

However, scaling neural networks to global atmospheric data introduces a fundamental geometric challenge: the Earth is a sphere. Standard convolutional neural networks (CNNs) assume a flat, Euclidean grid. Mapping a spherical surface to a flat plane introduces severe distortions and boundary discontinuities at the poles, leading to computational instability and artifacts during long-term simulations.

To resolve this, Bonev et al. developed Spherical Fourier Neural Operators (SFNOs). Instead of performing convolutions on a distorted flat grid, SFNOs leverage Spherical Harmonic Transforms to map atmospheric states into the spectral domain directly on the sphere.

[Global Atmospheric State] 
         │
         ▼  (Spherical Harmonic Transform)
[Spectral Domain Coefficients] 
         │
         ▼  (Linear Transform + Activation in Frequency Space)
[Updated Spectral Coefficients] 
         │
         ▼  (Inverse Spherical Harmonic Transform)
[Next Atmospheric State]

By learning the dynamics in the frequency domain rather than physical space, the model naturally handles global boundary conditions and eliminates polar singularities. This architectural adaptation ensures mathematical stability over extended rollouts, allowing researchers to simulate an entire year of climate dynamics in less than 13 minutes on a single GPU.

Hybrid Architectures and Differentiable Physics

While fully data-driven models excel at short- to medium-range weather forecasting, they can struggle with long-term climate projections. Purely statistical models are prone to "spectral degradation"—a phenomenon where forecasts become progressively smoother or "blurry" over time, failing to preserve small-scale kinetic energy. Furthermore, unconstrained neural networks can violate fundamental physical laws, such as the conservation of mass and energy.

To address these limitations, researchers are turning to hybrid architectures that blend numerical solvers with machine learning. A prominent example is NeuralGCM, developed by Kochkov et al.. This framework combines a coarse-grained, differentiable numerical core—which handles large-scale fluid dynamics using traditional physics equations—with neural network modules that predict sub-grid physics.

By keeping the large-scale fluid dynamics inside a physical solver, the model guarantees that mass and momentum conservation laws are strictly obeyed. Meanwhile, the neural components handle the complex, non-linear sub-grid interactions at a fraction of the cost of high-resolution numerical parameterizations. To combat spectral blurriness without full physical coupling, Zhdanov et al. introduced a Block-Sparse Attention mechanism that preserves spectral fidelity in purely data-driven models, matching the accuracy of high-resolution grids while running 150 times faster.

Downscaling and Massive Ensemble Generation

The massive speedups provided by neural emulators unlock two capabilities that were previously computationally impossible: ultra-high-resolution downscaling and massive ensemble forecasting.

Statistical Downscaling via Generative AI

Instead of running a global simulation at an ultra-fine resolution (e.g., 3 km), which requires millions of CPU hours, researchers can run a coarse, computationally cheap global model and use generative AI to upscale the outputs. For example, Perkins et al. developed HiRO-ACE, a framework that pairs a global atmospheric emulator with a diffusion-based downscaler. The diffusion model acts similarly to image super-resolution networks, injecting physically plausible high-frequency details (such as localized precipitation and tropical cyclone structures) into coarse data, providing a 100x speedup over explicit high-resolution modeling.

Sampling Rare Events with Huge Ensembles

Because weather is inherently chaotic, meteorologists run "ensembles"—multiple simulations with slightly perturbed initial conditions—to generate probabilistic forecasts. Traditional NWP models are limited by computational cost to roughly 50 or 100 ensemble members, which is insufficient to accurately sample low-probability, high-impact extreme weather events.

Using the computational efficiency of SFNOs, Cohen designed an ensemble forecasting system capable of generating thousands of calibrated members cheaply. In a follow-up study, Mahesh et al. leveraged this architecture to generate a massive 7,424-member ensemble. This unprecedented ensemble size drastically reduces sampling uncertainty, allowing scientists to identify and analyze rare, catastrophic weather anomalies well before they occur.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_computational_bottleneck_in_climate_mode_30046726,
  title = {computational bottleneck in climate modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/computational-bottleneck-in-climate-modeling/rdr_30046726e4b6450a89b7dbed67d90bbc},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_computational_bottleneck_in_climate_mode_30046726_v1,
  title = {computational bottleneck in climate modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/computational-bottleneck-in-climate-modeling/rdr_30046726e4b6450a89b7dbed67d90bbc/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake