TaiJi: State-Conditional Adaptive Combination of Machine-Learning and Physics-Based Global Weather Forecasts

Jiale Wang, Yining Li, Guihua Wang

arXiv preprint · 25 September 2026 · Fudan University (FuXi)

In one sentence. A learned CNN combines Pangu, GraphCast, FuXi and IFS forecasts with state-dependent weights, beating the best baseline on most variables out to 10 days.

Abstract

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 in nearly every combination of the eight core variables, lead times of up to 10 days and two skill metrics (RMSE and ACC), and this advantage extends to heavy-precipitation extremes. 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.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.