A data-driven approach to real-time reconstruction of turbulent flow fields above a coarse-grain bed using sparse observational boundary data

Reconstructing turbulent flow fields from sparse or partial measurements has become increasingly feasible with advances in data-driven machine learning. This study introduces a convolutional autoencoder-based neural network for reconstructing real-time velocity and turbulent kinetic energy (TKE) fields at the same sampling time as the boundary input over coarse-grain beds in a 100 mm × 100 mm region, solely using upstream boundary profiles. A scalable model architecture is established to handle various input/output data formats, and a physics-informed training strategy is implemented to enhance training efficiency and accuracy. Assessments demonstrate that the model effectively captures spatiotemporal flow patterns, particularly horizontal and total velocities, while time-averaged fields further reduce errors in real-time outputs. However, accuracy decreases for vertical velocity and high-TKE regions. Robustness tests indicate that the model generally remains stable under incomplete or corrupted inputs, whereas TKE reconstructions are more sensitive to those disturbances and require cleaner data and improved training. This study highlights the importance of embedding more physical laws within the model. Overall, the proposed model demonstrates promising potential for real-time flow field reconstruction that requires fewer input data, though further refinement is necessary to improve the generalizability, physical consistency, and application to more complex flow environments. • A convolutional autoencoder model is proposed for reconstructing 2D velocity and TKE fields using boundary measurements. • The physics-informed training strategy improves training efficiency, reconstruction fidelity, and model interpretability. • Sensitivity analyses are conducted to examine the model's robustness with incomplete and corrupted boundary data.