[Paper Review] Neural General Circulation Models for Weather and Climate

Dmitrii Kochkov, Janni Yuval+14 more|arXiv (Cornell University)|Nov 13, 2023

Meteorological Phenomena and Simulations24 citations

TL;DR

NeuralGCM is the first fully differentiable hybrid general circulation model of the Earth’s atmosphere, combining a neural-parameterized physics module with a differentiable dynamical core to match state-of-the-art weather and climate forecasts while enabling large compute savings.

ABSTRACT

General circulation models (GCMs) are the foundation of weather and climate prediction. GCMs are physics-based simulators which combine a numerical solver for large-scale dynamics with tuned representations for small-scale processes such as cloud formation. Recently, machine learning (ML) models trained on reanalysis data achieved comparable or better skill than GCMs for deterministic weather forecasting. However, these models have not demonstrated improved ensemble forecasts, or shown sufficient stability for long-term weather and climate simulations. Here we present the first GCM that combines a differentiable solver for atmospheric dynamics with ML components, and show that it can generate forecasts of deterministic weather, ensemble weather and climate on par with the best ML and physics-based methods. NeuralGCM is competitive with ML models for 1-10 day forecasts, and with the European Centre for Medium-Range Weather Forecasts ensemble prediction for 1-15 day forecasts. With prescribed sea surface temperature, NeuralGCM can accurately track climate metrics such as global mean temperature for multiple decades, and climate forecasts with 140 km resolution exhibit emergent phenomena such as realistic frequency and trajectories of tropical cyclones. For both weather and climate, our approach offers orders of magnitude computational savings over conventional GCMs. Our results show that end-to-end deep learning is compatible with tasks performed by conventional GCMs, and can enhance the large-scale physical simulations that are essential for understanding and predicting the Earth system.

Motivation & Objective

• Motivate replacing or augmenting traditional GCM parameterizations with learnable components to improve forecast skill and uncertainty quantification.
• Develop a fully differentiable hybrid model integrating a dynamical core with a neural physics module trained end-to-end.
• Enable accurate deterministic and ensemble forecasts across weather to climate timescales, including multi-day forecasts and decadal simulations.
• Demonstrate that the model can produce realistic climate features (seasonal cycles, tropical cyclones) and maintain stability over long integrations.
• Quantify computational efficiency and assess generalization and physical consistency relative to conventional models.

Proposed method

• Use a differentiable dynamical core that solves the hydrostatic primitive equations with moisture using a horizontal pseudo-spectral discretization and vertical sigma coordinates.
• Implement a single-column neural network to parameterize unresolved processes (clouds, radiation, precipitation) via an encode-process-decode architecture with weights shared across columns.
• Introduce encoders/decoders to map between ERA5 pressure-level data and model sigma-coordinate levels, including learned corrections to interfaces to reduce initialization shocks.
• Train end-to-end with rollout-based online training, gradually extending forecast horizons from hours up to 5 days, using losses that balance accuracy, sharpness, and bias in spectral and grid spaces.
• Incorporate stochasticity to generate ensemble forecasts via a CRPS-based loss and Gaussian-random-field noise with learned spatial/temporal correlations.
• Evaluate at multiple resolutions (2.8°, 1.4°, 0.7°) with 32 vertical levels, comparing deterministic and stochastic variants to ECMWF-ENS/HRES, GraphCast, and Pangu, and perform climate simulations with prescribed SST/SI.

Experimental results

Research questions

• RQ1 Can a fully differentiable hybrid GCM match state-of-the-art deterministic and ensemble weather forecasts across 1-15 day horizons?
• RQ2 Do end-to-end trained neural parameterizations, coupled with a differentiable dynamical core, produce stable long-term climate simulations with realistic emergent phenomena?
• RQ3 Are ensemble forecasts from NeuralGCM well-calibrated (CRPS, spread-skill) compared to traditional ensemble predictions?
• RQ4 What are the computational trade-offs and scalability gains when replacing traditional parameterizations with learned components?
• RQ5 Can NeuralGCM reproduce key climate features such as tropical cyclone tracks, Hadley circulation, monsoons, and seasonal cycles under AMIP-like setups?

Key findings

• NeuralGCM achieves competitive accuracy with best-in-class models for 1-10 day deterministic weather forecasts and 1-15 day ensemble forecasts.
• Stochastic NeuralGCM-ENS achieves lower ensemble-mean RMSE and CRPS than ECMWF-ENS across most variables and lead times, with a spread-skill ratio near one.
• At coarser resolutions (2.8°–1.4°), NeuralGCM can simulate climate behavior with realistic seasonal cycles, Hadley circulation, monsoons, and tropical cyclones under AMIP-like runs.
• In 40-year AMIP-like experiments with prescribed SST, NeuralGCM captures historical temperature trends and reduces temperature bias relative to some AMIP baselines, while maintaining realistic tropical warming structure.
• NeuralGCM offers substantial compute savings, operating at 8–40x coarser horizontal resolution than ECMWF IFS and cloud-resolving models, enabling 3–5 orders of magnitude faster simulations.
• Case studies show NeuralGCM-produced tropical cyclone tracks and frequencies are closer to ERA5 than some high-resolution baselines when regridded, illustrating realistic extreme-weather behavior.