Neural General Circulation Models

通用环流模型（GCMs）是天气和气候预测的基础。GCMs是基于物理的模拟器，将大尺度动力学的数值求解器与小尺度过程（如云形成）的调整表示相结合。最近，使用再分析数据训练的机器学习（ML）模型在确定性天气预报方面取得了与GCMs相当或更好的技能。然而，这些模型尚未展示出改进的集合预测能力，或者表现出足够长期天气和气候模拟的稳定性。在这里，我们提出了第一个将大气动力学的可微分求解器与ML组件相结合的GCM，并展示它可以生成与最佳ML和基于物理的方法相当的确定性天气、集合天气和气候预报。NeuralGCM在1-10天的预报中与ML模型竞争力相当，在1-15天的预报中与欧洲中期天气预报中心的集合预报相当。在预设海表温度的情况下，NeuralGCM可以准确跟踪多个十年的全球平均温度等气候指标，具有140公里分辨率的气候预报展现出像真实的热带气旋频率和轨迹等新兴现象。对于天气和气候，我们的方法比传统GCMs节省了数个数量级的计算成本。我们的结果表明，端到端的深度学习与传统GCMs执行的任务兼容，并且可以增强理解和预测地球系统所必需的大尺度物理模拟。

The paper aims to create a GCM that combines a differentiable solver for atmospheric dynamics with ML components, which can generate forecasts of deterministic weather, ensemble weather and climate on par with the best ML and physics-based methods while offering orders of magnitude computational savings over conventional GCMs.

The key idea of the paper is to use end-to-end deep learning to enhance the large-scale physical simulations that are essential for understanding and predicting the Earth system. The authors propose NeuralGCM, the first GCM that combines a differentiable solver for atmospheric dynamics with ML components, which achieves comparable or better skill than GCMs and ML models for deterministic weather forecasting and ensemble forecasts, and demonstrates sufficient stability for long-term weather and climate simulations.

The paper presents NeuralGCM, the first GCM that combines a differentiable solver for atmospheric dynamics with ML components, which can generate forecasts of deterministic weather, ensemble weather and climate on par with the best ML and physics-based methods while offering orders of magnitude computational savings over conventional GCMs. The authors show that 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. The paper also highlights the potential of end-to-end deep learning for enhancing large-scale physical simulations and understanding and predicting the Earth system.

Recent related work includes the use of machine learning models trained on reanalysis data for deterministic weather forecasting, which achieved comparable or better skill than GCMs, but have not demonstrated improved ensemble forecasts or sufficient stability for long-term weather and climate simulations. Some examples of related papers include 'Deep learning for weather forecasting: a comparison of deep neural networks and neural networks' and 'Machine learning for global weather and climate prediction'.