Data-Driven Post-Processing for High Fidelity Computational Fluid Dynamics

In today’s competitive aerospace industry, powerful computational tools are becoming increasingly accessible, thanks to significant advancements in GPU technology and the parallel processing capabilities provided by CUDA software. As a result, high-fidelity simulations are playing a growing role in industrial aerodynamic design, offering highly reliable solutions. However, the large volume of data generated by these simulations presents a pressing need for data reduction techniques and advanced post-processing tools that can keep pace with these developments, enabling comprehensive analysis of complex non-linear dynamical systems. This thesis aims to develop advanced data-driven post-processing techniques for high-fidelity simulations, specifically to improve the analysis of low-pressure turbine cascade performance. The work continues a research project initiated at Avio Aero (GE), which provided essential tools for its development. The post-processing procedure leverages Proper Orthogonal Decomposition (POD) to calculate and analyze profile loss in turbine cascades under low-speed conditions, focusing on turbulence and other unsteady phenomena related to wake-boundary layer interactions. The initial step involved calculating profile loss using standard post-processing methods. A Python code was then developed to extract data from the HiFi solver results and apply the data-driven algorithm. The reliability of the post-processing output was validated by comparison with standard techniques, and the advanced methods were employed to analyze unsteady loss. This approach enables the identification of coherent structures, the categorization and in-depth study of aerodynamic loss-generation phenomena, and ultimately supports designers in optimizing performance. The data-driven software developed enhances and supplements Computational Fluid Dynamics (CFD) prediction capabilities by constructing reduced-order models (ROMs) of the initial dataset. This allows for information compression, reduced disk memory usage, and low-cost predictions of aerodynamics at intermediate points of the training input.