在科学界的一次重大突破中，Google Research的研究团队成功开发了一种将传统基于物理建模与机器学习（ML）相结合的新方法——NeuralGCM，该方法在模拟地球大气层的准确性、计算效率和成本效益上都超越了现有模型。这一创新成果不仅为气象预测和气候研究提供了更强大的工具，而且有望在全球变暖和极端天气事件的预测方面发挥关键作用。
NeuralGCM通过将可微分的动力学核心与用于处理小规模过程的神经网络相结合，实现了高效率的模拟能力。动力学核心负责求解离散化的动力学控制方程，模拟在重力和科里奥利力作用下的大尺度流体运动和热力学过程。与此同时，学习物理模块则利用神经网络预测未解决过程对模拟场的影响，如云的形成、辐射传输、降水和亚网格尺度动力学。这种结合方式不仅提高了模型的准确性，还显著提升了计算效率，与当前最先进的模型相比，NeuralGCM在相似或更高准确度下的计算效率提高了3到5个数量级。
这一创新成果的发布标志着人工智能在气象和气候建模领域的重要进展。NeuralGCM不仅能够生成2至15天的天气预报，而且在1至10天的预报精度上与机器学习模型相当，而在1至15天的预报上则与欧洲中期天气预报中心的集合预报相媲美。NeuralGCM的模型结构设计和端到端训练方法使其能够在多个时间步骤上在线调整耦合系统的行为，从而克服了以往使用ML增强气候模型时在数值稳定性方面的挑战。
NeuralGCM的成功不仅展示了人工智能在气象和气候研究中的巨大潜力，也为未来更准确、更高效地预测全球变暖和极端天气事件提供了新的工具。随着这一成果的发布，科学家们对理解地球气候系统、预测气候变化趋势以及采取有效应对措施以减轻全球变暖的影响有了更强的信心。这一突破性的进展标志着科学界在利用先进计算技术解决全球性环境问题方面迈出了重要一步。
英语如下：
News Title: “Google’s New Model: More Accurate, Faster, and Affordable Weather Forecasting”
Keywords: NeuralGCM, Google Research, Weather Prediction
Content: In a significant scientific breakthrough, the Google Research team has successfully developed NeuralGCM, a novel approach that merges traditional physics-based modeling with machine learning (ML). This method surpasses existing models in terms of accuracy, computational efficiency, and cost-effectiveness when simulating Earth’s atmosphere. The innovation not only provides a powerful tool for meteorological predictions and climate research but also promises to play a crucial role in forecasting global warming and extreme weather events.
NeuralGCM combines a differentiable dynamical core with neural networks to handle small-scale processes, enabling high-efficiency simulations. The dynamical core solves the discretized dynamical control equations to simulate large-scale fluid motions and thermodynamic processes under the influence of gravity and Coriolis force. Meanwhile, the physics-informed module uses neural networks to predict the impact of unresolved processes on the simulated fields, such as cloud formation, radiation transfer, precipitation, and sub-grid-scale dynamics. This integration not only enhances model accuracy but also significantly boosts computational efficiency, with NeuralGCM achieving 3 to 5 orders of magnitude higher efficiency than the current state-of-the-art models at similar or higher accuracy levels.
The release of this innovation marks a significant advance in artificial intelligence’s application in meteorological and climate modeling. NeuralGCM can generate weather forecasts from 2 to 15 days ahead, matching the precision of machine learning models for 1 to 10 days ahead and rivaling the European Centre for Medium-Range Weather Forecasts’ ensemble forecasts for 1 to 15 days ahead. The design of NeuralGCM’s model structure and end-to-end training method allows for online adjustments of the coupled system’s behavior across multiple time steps, overcoming challenges in numerical stability that arise when using ML to enhance climate models.
The success of NeuralGCM showcases the vast potential of AI in meteorological and climate research and provides new tools for more accurate and efficient predictions of global warming and extreme weather events. With the release of this breakthrough, scientists are gaining stronger confidence in understanding Earth’s climate system, forecasting climate change trends, and taking effective measures to mitigate the impacts of global warming. This landmark achievement signifies a significant step forward for the scientific community in leveraging advanced computing technologies to address global environmental challenges.
【来源】https://www.jiqizhixin.com/articles/2024-07-23-7
Views: 1