With its "Flightpath 2050" strategy, the European Commission has outlined a framework for the aviation industry that aims to reduce emissions as well as fuel and energy consumption. Among other things, this requires more efficient engines. In the ARIADNE project, an interdisciplinary team at Graz University of Technology (TU Graz) has created the basis for achieving the desired efficiency gains more quickly. To this end, the researchers have combined years of flow data on intermediate turbine ducts with AI and machine learning and developed a model that much more quickly and efficiently tests the impact of changes to a wide range of geometry parameters on efficiency.
Intermediate turbine ducts offer a lot of optimisation potential
"Intermediate turbine ducts are an essential component of aircraft engines," says project manager Wolfgang Sanz from the Institute of Thermal Turbomachinery and Machine Dynamics at TU Graz. "They guide the flow between the high-pressure and low-pressure turbines, which run at different speeds. However, these intermediate ducts are quite heavy, which is why they need to be as short, small and light as possible while still achieving high levels of efficiency. There is still a lot of potential for optimisation here."
Based on its own research in collaboration with renowned aircraft engine manufacturers, the institute has built up an extensive database of measurement data and flow simulations. In order to utilize this reservoir of information to optimise components and entire engines, Wolfgang Sanz and doctoral student Marian Staggl collaborated with Franz Wotawa's research group at the Institute of Software Engineering and Artificial Intelligence at TU Graz as well as two corporate partners. Together, they pursued three different AI-supported approaches.
Success with reduced order modelling
Reduced order models proved to be the most successful. These models search for similarities in the data and use only the most significant common features for simulation. This leads to an enormous acceleration of the calculations, which run several orders of magnitude faster than a complete flow simulation. Although these models can entail certain losses in accuracy, they allow to predict trends and to identify optimisation potential by linking them with the simulation. Another advantage of the independently developed model was the ability to quickly recognise changes in efficiency when a parameter, such as the length of the transition duct, changes.
In contrast, surrogate models had certain limitations, as they are mainly based on interpolation of existing data. Outside the validated flow data range, the results were inaccurate because the database was too small. PINNs (Physics Informed Neural Networks), which attempt to integrate physical differential equations into a neural network, were also investigated as part of the project. However, further developments are still required before they can be used in practice.
Extension to three-dimensional simulations
The research team is already planning the next steps, as the reduced order model has so far only modelled the intermediate turbine ducts in two dimensions. The extensive database on turbine ducts and the reduced order model created in the project will be made available online to other research groups, allowing them to work on a three-dimensional simulation model similar to the team at TU Graz. For Wolfgang Sanz, however, working with machine learning has already opened up new approaches. "From the results of the machine learning approaches, we were able to recognise dependencies and trends that we would never have thought of otherwise."
