Neural general circulation models for weather and climate

Neural general circulation models (GCMs) are introduced as a novel approach for weather and climate prediction, combining a differentiable solver for atmospheric dynamics with machine-learning components. These models aim to provide accurate forecasts for deterministic weather, ensemble weather, and climate simulations. The NeuralGCM demonstrates competitive performance with machine-learning models and traditional GCMs for various forecast timescales, offering computational efficiency and stability. By training on historical data and integrating machine learning with physical simulations, NeuralGCM shows promise in accurately tracking climate metrics and simulating emergent phenomena like tropical cyclones. The model's differentiable dynamical core allows for end-to-end training, enabling interactions between learned physics and large-scale dynamics. NeuralGCM's performance is evaluated against conventional physics-based models and machine-learning approaches, showcasing its accuracy and physical consistency in weather forecasting and climate simulation. The hybrid nature of NeuralGCM combines the interpretability and successful track record of traditional GCMs with the efficiency and accuracy of machine learning, addressing limitations of existing models and offering a promising avenue for advancing weather and climate prediction capabilities.