Deep-learning-based particle inertia bias corrector for shock-dominated PIV data

Abstract Tracer particle inertia introduces systematic bias in supersonic particle image velocimetry (PIV), degrading the accuracy of measured velocity fields, particularly across shock waves. This study presents a physics-aware deep neural network approach for correcting particle inertia bias directly from raw PIV image pairs. The proposed model, termed the Bilateral Inertia-Correction Sparse Neural Network (BICSNet), employs a bilateral convolutional encoder–decoder architecture conditioned on freestream Mach number and freestream Reynolds number. BICSNet is trained using a large synthetic dataset designed to emulate realistic PIV imaging of oblique shocks across a range of Mach numbers, shock strengths, deflection angles, particle properties, and image parameters. The model achieves substantial reductions in particle inertia bias for in-distribution flow conditions, with predictions that closely match analytical solutions. After appropriate intensity-domain preprocessing, the methodology is applied to experimental PIV measurements of shock interactions. Significant improvements in velocity gradient recovery are observed across isolated and interacting shocks, demonstrating the applicability of the proposed framework to real-world experimental workflows.