发表机构
Fudan University（复旦大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
TaiJi提出一种轻量级卷积神经网络，基于近期大气状态自适应组合ML与NWP预报，在WeatherBench 2上以极低训练成本实现全面最优性能。
AI 中文摘要
数据驱动的机器学习天气模型现在可与业务数值天气预报（NWP）相媲美，在某些方面甚至超越，但没有任何单一模型能在变量、气压层、提前时间或区域上占据主导地位，而且开发日益庞大的个体模型的边际收益正在递减。我们提出TaiJi（取自中国阴阳和谐的概念，即对立面的互补性），一个时空自适应集成框架，将最优模型权重组合重新表述为一个学习预测问题：一个轻量级卷积神经网络，以近期大气状态为条件，在每个网格点和提前时间预测一组仿射组合权重和组成预报的加性残差。该组合器在Pangu-Weather、GraphCast、FuXi、IFS-HRES和IFS-ENS集合平均上端到端训练，这是一个明确混合的ML和NWP组成集合，在两块消费级RTX 4090 GPU上约8小时完成，训练成本至少比其任何ML组成部分低一个数量级。在WeatherBench 2协议（2020年测试年）下，TaiJi在八个核心变量和最长10天的提前时间上，在RMSE和ACC指标下，均优于每个变量和提前时间的最强基线，胜率接近100%，且这一优势在极端场景（包括热带气旋路径和强降水）中得以保持。我们进一步表明，学习到的权重随纬度、季节和提前时间连贯变化，而非遵循单一固定规则。因此，TaiJi提供了一条计算成本低廉的途径来提高全球预报技能，它建立在现有预报模型的大量投资之上，而非取代它们。
英文摘要
Data-driven machine-learning weather models now rival, and in some respects surpass, operational numerical weather prediction (NWP), yet no single model dominates across variables, pressure levels, lead times, or regions, and the marginal returns from developing ever-larger individual models are diminishing. We present TaiJi (after the Chinese concept of yin-yang harmony, the complementarity of opposites), a spatiotemporal adaptive ensemble framework that recasts optimal model-weight combination as a learned prediction problem: a lightweight convolutional neural network, conditioned on the recent atmospheric state, predicts at every grid point and lead time a set of affine combination weights and an additive residual for the constituent forecasts. The combiner is trained end-to-end over Pangu-Weather, GraphCast, FuXi, IFS-HRES, and the IFS-ENS ensemble mean, an explicitly hybrid set of ML and NWP constituents, on two consumer-grade RTX 4090 GPUs in about 8 hours, at least an order of magnitude below the training cost of any of its ML constituents. Under the WeatherBench 2 protocol (2020 test year), TaiJi outperforms the strongest baseline for each variable and lead time across the eight core variables and lead times of up to 10 days, under both RMSE and ACC, with a near-100% win rate, and this advantage is preserved across extreme scenarios, including tropical-cyclone tracks and heavy precipitation. We further show that the learned weights vary coherently with latitude, season and lead time rather than following a single fixed rule. TaiJi thus offers a computationally inexpensive route to improving global forecast skill that builds on, rather than replaces, the substantial investment already made in existing forecast models.
