Session: 32-03 CFD Studies

Paper Number: 153885

The Effect of Training Data on Predicting Turbulent Flow Through a Linear Cascade Using Physics-Informed Neural Networks 

This paper investigates the effect of training data on the accuracy of turbulent flow predictions through a linear turbine cascade using physics-informed neural networks (PINNs). While it is well known that PINNs can solve the unclosed Reynolds-averaged Navier-Stokes (RANS) equations when sufficient training data are available, the specific characteristics of the data – such as the quantity, location, and type – required for accurate predictions remain largely uncertain but are believed to be application-specific. To explore this, a PINN is constructed to solve the RANS equations leveraging training data from a time-averaged Large-Eddy Simulation (LES) simulation of transonic flow around a turbine blade. The training data are then selectively sampled to reflect what can be measured in a linear turbine cascade experimental facility. This sampling includes blade loading (pressure distribution) and data from computational survey planes at various axial locations downstream of the blade. For each survey, the PINN is trained on the time-averaged LES-solution data of the velocity components, temperature, and pressure, while also varying the Reynolds stresses and turbulent heat fluxes to study which quantities are essential for accurate predictions. The objective is to provide guidance on which quantities should be prioritized to achieve high-fidelity predictions with a PINN, using the minimum amount of data possible. 

Presenting Author: Ezra McNichols NASA

Presenting Author Biography: Ezra McNichols is a research aerospace engineer at NASA Glenn Research Center and a PhD student at The Ohio State University. He currently serves as the Turbomachinery Technical Lead for the Advanced Air Transport Technology (AATT) project. His research focuses on heat transfer improvement in heat exchangers, wicked and oscillating heat pipes, aerodynamics and heat transfer in turbomachinery, and applying machine learning methods for design and analysis.

Authors:

Ezra McNichols NASA
Jeffrey Bons The Ohio State University