Self-supervised temporal reconstruction of reanalysis fields at subhourly scales

Abstract Fast‑evolving weather events typically develop on 5–15 min scales and can significantly influence human activities and planning. However, most global atmospheric reanalyses provide data only every 1–3 hours, making it difficult to capture rapid changes such as short-duration heavy rainfall. Traditional machine learning approaches rely on high-frequency data as supervision labels during training, which limits their use when such data are unavailable. Here, we address this limitation by introducing a self-supervised framework that reconstructs subhourly details from long-term, low-resolution climate records. We observe that key atmospheric variables show scale-invariant patterns over time, without a dominant frequency and following a power-law decay. Therefore, we train the model using three tasks—extrapolation, interpolation, and backcasting—to help it generalize from hours to minutes. Compared with deep learning approaches, our method gives better results in cylinder-flow experiments. When applied to climate data—such as humidity, temperature and precipitation—it provides more accurate 15-min reconstructions across regions than the WRF model. Using NOAA-20CRv3 for training, the model gains the ability to reconstruct a global 15-min record dating back to 1806, facilitating detailed historical climate analyses.