Weather forecasting is crucial for scientific research and society. Recently, deep learning (DL) methods have achieved significant advancements in medium-range weather forecasting. However, they generally depend on the initial fields generated by the computationally expensive four-dimensional variational (4DVar) data assimilation (DA) technique, which limits their real-time applicability in multivariate three-dimensional (3D) weather forecasting. Here we propose 4DVarFormer by exploring the potential of integrating the 4DVar constraint into an attention-based neural network. 4DVarFormer eliminates the need for background error covariance statistics and the complex adjoint model development. It can generate multivariate 3D weather states within 0.37 s. Moreover, 4DVarFormer can capture inter-variable relationships, allowing the assimilation of observed variables to correct unobserved variables. Hence, medium-range forecasts initiated by 4DVarFormer outperform those of DL-based DA methods and achieve performance comparable to the forecasts initiated by ERA5 reanalyses. These promising findings contribute to future advancements in integrated end-to-end DL weather forecasting systems.

天气预报对于科学研究和社会至关重要。最近，深度学习（DL）方法在中期天气预报方面取得了重大进展。然而，它们通常依赖于计算成本昂贵的四维变分（4DVar）数据同化（DA）技术生成的初始场，这限制了它们在多元三维（3D）天气预报中的实时适用性。在这里，我们通过探索将 4DVar 约束集成到基于注意力的神经网络中的潜力，提出了 4DVarFormer。 4DVarFormer 消除了背景误差协方差统计和复杂的伴随模型开发的需要。它可以在 0.37 秒内生成多元 3D 天气状态。此外，4DVarFormer 可以捕获变量之间的关系，允许同化观察到的变量来纠正未观察到的变量。因此，4DVarFormer 发起的中期预测优于基于 DL 的 DA 方法，并且达到与 ERA5 重新分析发起的预测相当的性能。这些有希望的发现有助于集成端到端深度学习天气预报系统的未来进步。