Neural General Circulation Models for Weather and Climate (2311.07222v3)

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, 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.

Analysis of Neural General Circulation Models for Weather and Climate

The paper "Neural General Circulation Models for Weather and Climate" introduces NeuralGCM, an innovative approach that combines traditional General Circulation Models (GCMs) with ML techniques for atmospheric modeling. The model is a pioneering differentiable hybrid that uses a fully-differentiable solver for atmospheric dynamics, integrating learning components directly into its framework. This method draws on strengths from both physics-based simulations and data-driven techniques, aiming to enhance the reliability, efficiency, and breadth of atmospheric models.

Core Contributions and Methodology

Central to this work is the ability of NeuralGCM to forecast deterministic weather, ensemble weather, and climate states with a level of precision on par with or surpassing existing models. Traditional GCMs are heavily reliant on physical parameterizations, and while mature, these models exhibit persistent biases and require significant computational resources. Recent advancements in ML offer competitive deterministic weather forecasting capabilities but suffer from instability in ensemble predictions and long-term simulations.

The authors address these issues by integrating machine learning components with physics-based models, creating a fully differentiable GCM that maintains stability over extended simulations. NeuralGCM's architecture features:

1. Differentiable Dynamics: Implemented in JAX, NeuralGCM's dynamical core solves the hydrostatic primitive equations using spectral methods on a sigma-coordinate system. This setup enables efficient time-stepping and tackles issues related to rapid oscillations post-initialization.
2. Learned Physics Module: This component augments a pseudo-spectral dynamical core with a neural network-based representation for subgrid processes. These include cloud formation and radiative transport, which are traditionally handled by cumbersome parameterizations.
3. End-to-End Learning: The model employs a training strategy that incrementally increases forecast horizons. This end-to-end approach leverages automatic differentiation to fine-tune the interaction between parameterized physics and resolved dynamics.

Evaluation and Results

NeuralGCM was extensively evaluated across weather forecasting and climate simulation scenarios. It demonstrated competitive skill with current state-of-the-art models in:

• Medium-range Weather Forecasting: NeuralGCM performs well in short-to-medium-range deterministic forecasts (1-10 days) and ensemble forecasts (1-15 days), where it outperforms traditional and ML models in several metrics, notably CRPS.
• Seasonal and Decadal Climate Simulations: The model can execute stable multi-decade climate projections with realistic emergent phenomena, such as tropical cyclones, contributing to its verification as an effective tool for climate science.

NeuralGCM also presents computational advantages over traditional GCMs, offering substantial savings in computing resources, enabling practical long-term simulations, and handling stochastic elements effectively with ensemble prediction techniques.

Implications and Future Directions

The implications of this research are significant both theoretically and practically. By successfully integrating ML within the structure of a GCM, the paper opens new pathways for developing hybrid models that leverage data-driven capabilities while retaining the rigor of physical laws. This could lead to improvements in ensemble forecasting, better representation of atmospheric processes, and reduction in computational costs for climate models.

Theoretically, NeuralGCM embodies the hypothesis that training models on short-term predictive tasks can enhance their long-term climate projection abilities. This approach can serve as a framework to optimize models based on available data and observational comparisons, ensuring realism in forecasts and simulations.

Future work should focus on enhancing model generalization, particularly under climate change scenarios where observed conditions may diverge significantly from historical data. Extending this hybrid modeling approach to other Earth system components, such as land and ocean interfaces, will further augment its utility in comprehensive climate system modeling.

In conclusion, NeuralGCM represents a significant advancement in the development of hybrid models for weather and climate prediction, offering a promising direction for future research and operational forecasting systems.