Image-Based Prediction of Lift and Drag Coefficients of a Circular Cylinder Using Convolutional Neural Networks and Flow Visualization

Force measurement is a fundamental diagnostic technique in experimental fluid mechanics. However, conventional instrumentation, such as mechanical balances and pressure taps, is often intrusive, geometry-dependent, and mechanically constrained. This study proposes a non-contact, image-driven force-estimation framework wherein convolutional neural networks (CNNs) infer the instantaneous lift and drag coefficients of a circular cylinder directly from wake streakline flow-visualization images. A comprehensive hybrid numerical–experimental dataset was established across laminar and turbulent regimes. Navier–Stokes simulations (DNS at Re=150; LES at Re=500 and 3900) were post-processed into synchronized streakline–force pairs, complemented by water-tunnel dye-visualization experiments at Re=500, 8000, and 12,000. Five CNN architectures, including lightweight custom networks and vision backbones (AlexNet, ResNet18), were trained and validated. For numerical datasets, all models show high prediction accuracy in lift and drag coefficients. In simulation-to-experiment transfer at Re=500, the models capture the dominant vortex-shedding Strouhal number within 12.0% (St=0.233 vs. 0.208), while CNNModel_v5 predicts the mean drag coefficient within 0.35%. On experimental datasets, ResNet18 achieved the best predictive performance among the models tested under random splitting. These results indicate that wake flow-visualization images contain rich physical representations for non-intrusive fluid force estimation.