Neural network based on high freedom dynamic collaboration of flows for meteorological data temporal downscaling

Temporal refinement of meteorological fields is a critical yet challenging task, as it requires reconstructing physically consistent intermediate states from sparsely sampled observations. Conventional approaches either rely on computationally intensive numerical simulations or exhibit limited generalization when applied to complex atmospheric conditions. Recent advances in deep learning-based frame interpolation offer a data-driven alternative for modeling temporal evolution, providing new opportunities for meteorological time downscaling. This study proposes a deep neural network specifically designed to perform temporal downscaling of multivariate meteorological fields. The model independently estimates spatially adaptive kernel parameters for each variable and location, while dynamically integrating information across multiple meteorological fields. It generates output frames under the guidance of the feature space. Compared with existing approaches, the proposed framework demonstrates enhanced flexibility in capturing complex and heterogeneous atmospheric dynamics. Experiments conducted on 2 m temperature, surface pressure, specific humidity at 1000 hPa, and total precipitation demonstrate that the proposed method achieves high accuracy, robustness, computational efficiency, strong scalability and transferability.