This thesis investigates the potential for Physics-Informed Neural Networks (PINNs) to
reconstruct dense, 3D flow fields from sparse experimental wind tunnel data over the Gaussian
Boeing bump geometry. The difficulty in experimentally obtaining flow field data that is
This thesis investigates the potential for Physics-Informed Neural Networks (PINNs) to
reconstruct dense, 3D flow fields from sparse experimental wind tunnel data over the Gaussian
Boeing bump geometry. The difficulty in experimentally obtaining flow field data that is
simultaneously accurate, 3-dimensional and offering a wide field of view makes PINNs an
intriguing tool for flow enhancement due to their ease of data-acquisition.
This work attempts to reconstruct 3D separated velocity fields from 2D, two-component PIV
and static pressure data at Reynolds numbers up to 3 million. The predicted flow field is
validated with 2D, three-component PIV data along the untrained axis, assessing the
reconstruction accuracy. The various data types used in training the PINN are assessed for
their contributions to a successful reconstruction of the flow field. The PINN method's inherent
approximation of velocity gradients over the domain also allows for the prediction of surface
shear stresses. The PINN's capacity for reconstructing a variety of parameters is only curtailed
by the accuracy of its predictions, which remain costly and data-driven.
