An open-source, physics-aware, data-driven tool for automated particle image velocimetry post-processing

Abstract

Particle image velocimetry (PIV) is a powerful diagnostic tool for analyzing the properties and structures of laminar and turbulent flows. Modern PIV systems provide high spatial and temporal resolution measurements, often generating hundreds of gigabytes of raw data. Before these data can be used for flow analysis, they must be post-processed to identify and remove erroneous velocity vectors. Post-processing techniques typically require substantial manual intervention by highly trained analysts, making the process tedious, time-consuming, and inherently subjective. To address these challenges, we developed a physics-aware, data-driven framework for automated PIV post-processing. Our machine learning-based tool, PIVPostProcess, employs an ensemble of gradient-boosted decision trees trained on real experimental PIV datasets and is validated using direct numerical simulation data with synthetic noise. The fully trained model achieves excellent performance, with an F1 score—a performance metric that combines precision and recall—of 0.991 on the testing dataset. To evaluate generalization beyond the training distribution, we apply the model to six additional experimental datasets featuring previously unseen flow fields. The model achieves a mean F1 score of 0.990 across tested experimental datasets, demonstrating robust predictive capability. Importantly, this automated technique processes each dataset significantly faster compared to manual post-processing. By substantially reducing processing time while maintaining high accuracy, this tool accelerates and standardizes PIV vector validation approaches.