Exploiting temporal coherence for self-supervised denoising of turbulent flows

Denoising plays a crucial role in experimental fluid mechanics, where measurement noise can significantly contaminate flow structures and distort statistical quantities. A self-supervised denoising framework, incorporating a squeeze-and-excitation (SE) attention module and a convolutional long short-term memory (ConvLSTM) unit, referred to as SE–ConvLSTM, is developed by exploiting the temporal coherence of flows. The model reconstructs the target flow field from a sequence of preceding noisy snapshots, with performance primarily governed by the temporal correlation of the input data, determined by the input time span and the temporal interval between snapshots. The denoising capability of SE–ConvLSTM is systematically evaluated for two-dimensional flow fields under a wide range of noise intensities, noise types, and flow configurations, with comparisons against both classical and deep-learning methods. The results demonstrate that SE–ConvLSTM achieves superior performance for temporally uncorrelated noise, enabling more accurate noise removal while preserving fine-scale flow structures, spatial continuity, and turbulent statistics, particularly under severe noise contamination. The denoising performance is closely related to the ratio of the integral time scales of the noisy and reference flows, with improved accuracy obtained when temporal coherence is better preserved in the noisy field. Extension to three-dimensional flow fields further confirms the robustness of the framework in reconstructing volumetric structures. Overall, the SE–ConvLSTM provides an effective and generalizable approach for denoising noisy flow measurements.