Characterization of Turbulent Kinetic Energy and Shear Layers in Sudden Expansion Flows: A Physics-Informed Machine Learning Approach.

PURPOSE: Traditional Computational Fluid Dynamics (CFD) methods used in the hydrodynamic design of blood-contacting medical devices create a bottleneck in optimization processes due to their heavy dependence on mesh structure and high computational costs. METHODS: In this study, using a mesh-free Physics-Informed Neural Network (PINN) approach, complex flow fields and Turbulent Kinetic Energy (TKE) distributions in FDA reference nozzle geometries were mapped with high resolution. Unlike artificial intelligence algorithms that merely fit data, the proposed model directly integrates the axisymmetric mass conservation equation into the loss function as a mathematical constraint. The model was trained on experimental PIV/LDV data over a wide range covering laminar, transitional, and fully turbulent regimes ( R e = 500 - 6500 ). RESULTS: The model predicted sharp velocity gradients in the shear layers with high accuracy. Additionally, a practical empirical correlation for peak TKE was derived within the investigated conditions, serving as a rapid hydrodynamic indicator. CONCLUSION: The findings demonstrate that this physics-constrained approach reduces the need for mesh generation and offers a computationally lightweight framework for the continuity-constrained reconstruction and regularization of complex and inherently noisy experimental flow fields.