Breaking the computational barrier in high-resolution weather forecasting with decoupled training

Achieving high-resolution global AI weather forecasts remains computationally prohibitive, constrained by extensive GPU memory demands and prolonged training times. Existing approaches impose trade-offs—balancing speed, memory, or numerical precision. To circumvent these trade-offs, we introduce DeTI, a Decoupled Training–Inference framework that trains exclusively on high-resolution regional patches while incorporating global context from a pretrained, frozen low-resolution model. DeTI reduces computational costs yet preserves fidelity. Remarkably, it generalizes seamlessly during inference, enabling the extension of regional learning to global forecasts. A 0.25° model trained in only two days on eight NVIDIA A100 GPUs attains accuracy comparable to state-of-the-art systems. Moreover, DeTI scales efficiently to 0.09° resolution, demonstrating superior capability in high-impact scenarios, excelling in predicting typhoon intensity and capturing the fine-grained structure of heatwaves, paving the way for future kilometer-scale AI weather models. 针对全球高分辨率AI天气预报显存消耗大,训练周期长等瓶颈, 本文提出DeTI(解耦训练-推理)框架.该框架打破了训练与推理分辨率必须一致的限制, 采用“区域高分辨率训练”辅以“全球低分辨率背景引导”策略, 实现了计算开销的有效解耦.实验证实, 仅需8块A100 GPU训练2天, 0.25°模型精度即可媲美主流大模型.此外, DeTI可扩展至0.09°分辨率, 在台风强度预测及热浪精细表征中表现优异, 为开发公里级AI天气模型提供了高效路径.