Neural general circulation models for modeling precipitation

Janni Yuval, Ian Langmore, Dmitrii Kochkov, Stephan HoyerScience 2026 · DOI

TL;DR

This paper introduces a novel hybrid climate model, NeuralGCM, that directly trains on satellite precipitation data. This approach significantly improves precipitation simulation accuracy for both weather forecasting and long-term climate projections, outperforming traditional GCMs and reanalysis datasets while being orders of magnitude faster.

What Problem Does This Paper Address?

General Circulation Models (GCMs), the workhorses of climate science, notoriously struggle to accurately simulate precipitation. This is a critical issue because precipitation is fundamental to weather, climate, and human activities, especially extreme events and the diurnal cycle. While hybrid models combining physics and machine learning (ML) have shown promise, they've historically relied on high-fidelity simulations for training, inheriting their biases. This paper tackles the challenge of improving precipitation simulation in GCMs by developing a hybrid model that can be trained directly on observational data, bypassing the limitations of relying solely on simulated data.

Methods

The core innovation lies in a hybrid model built upon the differentiable NeuralGCM framework. This framework allows for end-to-end (online) training, meaning the entire model, including its physics and ML components, can be optimized simultaneously. The key advancement here is training the ML component (a neural network parameterization) directly on satellite-based precipitation observations (IMERG) rather than solely on outputs from other complex models.

The model consists of two main components:

  • A differentiable dynamical core: This handles the fundamental physics of atmospheric motion.
  • A learned physics module: This is a neural network that parameterizes processes not resolved by the dynamical core, specifically focusing on precipitation.

To ensure physical consistency, the model predicts precipitation and then diagnoses evaporation by enforcing the atmospheric water budget. This is a crucial modification from previous NeuralGCM versions that only predicted net precipitation minus evaporation. The training process involves minimizing a loss function that compares model predictions to ground truth data, using the Continuous Ranked Probability Score (CRPS) for weather forecasting and other metrics for climate simulations.

Datasets

  • Training Data:
    • ERA5: Used for atmospheric variables and evaporation.
  • IMERG V07 "final" dataset: Used for precipitation. This is a key dataset as the model is trained directly on it.
  • Evaluation Data:
    • IMERG: Used as a primary ground truth for precipitation.
  • Global Precipitation Climatology Project (GPCP): Used as an independent observation-based dataset for precipitation.
  • ERA5 Reanalysis: Used for comparison.
  • CMIP6 Models: Various AMIP and historical runs from CMIP6 are used for comparison.
  • GFDL's X-SHIELD: A global cloud-resolving model used for comparison.
  • ECMWF Ensemble (ENS): Used for medium-range weather forecasting comparisons.

A notable aspect is the direct use of satellite observations (IMERG) for training, which differs from previous hybrid models that relied on outputs from high-resolution simulations.

Key Insights and Ideas

  • Direct training on observations is powerful: By training directly on satellite precipitation data (IMERG), the model achieves significantly better precipitation simulation than models trained on other simulations. This highlights the value of observational data in improving climate model fidelity.
  • Differentiable framework enables end-to-end learning: The differentiable nature of NeuralGCM is crucial, allowing the entire model to be optimized simultaneously, leading to more coherent and accurate predictions.
  • Water budget constraint is vital: Explicitly modeling precipitation and diagnosing evaporation via the water budget leads to more physically consistent results compared to models that only predict net P-E.
  • Hybrid models can outperform traditional GCMs: This work demonstrates that a hybrid ML-physics model, even at a relatively coarse resolution (2.8°), can surpass established GCMs and reanalysis data for precipitation simulation.
  • Speed matters: The model is orders of magnitude faster than traditional GCMs, enabling larger ensembles and more extensive simulations.

Key Results

  • Weather Forecasting: NeuralGCM outperforms the ECMWF ensemble (ENS) across all 15 forecast days in terms of CRPS, RMSE, spread-skill ratio, and Brier score when evaluated against IMERG.
  • Climate Simulations:
    • Mean Precipitation: NeuralGCM shows a 40% reduction in global mean absolute error (MAE) compared to CMIP6 AMIP runs, achieving an MAE comparable to ERA5 despite running freely (not assimilated).
  • Precipitation Extremes: NeuralGCM more accurately reproduces the frequency distribution of precipitation rates and the spatial patterns of annual maximum daily precipitation (Rx1day) compared to ERA5 and CMIP6 models.
  • Diurnal Cycle: NeuralGCM captures the timing of peak diurnal precipitation more accurately than ERA5 and GFDL AMIP runs, particularly over land, shifting the peak to the late afternoon/evening as observed.
  • Computational Efficiency: The model simulates approximately 1200 simulated years per day on a single TPU, significantly faster than traditional GCMs.

Limitations and Future Work

  • Resolution: The current 2.8° resolution limits immediate practical applications for high-resolution weather forecasting. Future work could explore statistical downscaling or higher-resolution models.
  • Stability: While improved, the stable model was obtained by training multiple models and selecting the most stable one. Further research is needed to understand and address factors influencing model stability.
  • Conflicting Datasets: Inconsistencies between ERA5 and IMERG necessitated careful tuning of the loss function. Developing unified datasets for training would be beneficial.
  • NN Parameterization: The study used an NN to parameterize all unresolved processes. Future work could explore coupling the differentiable core with traditional parameterizations or optimizing their free parameters.
  • Diurnal Oscillations: The precipitation module exhibits some unrealistic oscillations in intensity during the first day of forecasting, though this is not present in the "NeuralGCM-evap" variant.

Summary updated

Abstract

General circulation models (GCMs) struggle to accurately simulate precipitation, particularly extremes and the diurnal cycle, which are crucial for both human activities and natural processes. Although hybrid models combining machine learning and physics offer a promising avenue to improve the simulation of precipitation, they have yet to outperform existing GCMs. Here, we present a hybrid model built on the differentiable NeuralGCM framework. This differentiability facilitates direct training on satellite-based precipitation observations, unlike previous attempts at hybrid models that relied on high-resolution simulations as training data. Our model runs at 2.8° resolution and, in the context of climate, demonstrates substantial improvements over existing GCMs, the ERA5 reanalysis, and a global cloud-resolving model in simulating precipitation. In the context of mid-range precipitation forecasting, it outperforms the ECMWF ensemble. This advance paves the way for more reliable simulations of current climate and demonstrates how training on observations can be used to improve GCMs.
BibTeX
@inproceedings{yuval2026neural,
  author = {Janni Yuval and Ian Langmore and Dmitrii Kochkov and Stephan Hoyer},
  title = {{Neural general circulation models for modeling precipitation}},
  booktitle = {Science},
  year = {2026},
  doi = {10.1126/sciadv.adv6891},
}
Figure