The past decade has seen great strides within aeronautics. There has been a sustained effort towards electric propulsion, the deployment of more fuel-efficient aircrafts for commercial travel, and even new demand for urban aerial taxis. Across all these scales - large commercial aircrafts, medium sized jets and smaller vertical take-off and landing (VTOL) vehicles - safety, reliability, and efficiency continue to remain paramount. This is especially important given the recent string of aviation incidences involving fatalities. To ensure both existing machinery and new aviation concepts comply to stringent standards, there is a strong need for more physically representative digital twins of aircraft systems.
Explaining the science
Digital twins are computational representations of aeronautic assets that can be used to model, optimise and predict asset performance.
More accurate digital twins permit engineers to flag potential issues, overhaul components and even re-design components well before any flight risks emerge. However, to usher in a new era of digital characterisation of such complex systems, a deeper understanding of underlying aerothermodynamics and aeroelasticity has to be paired with machine learning methods, ensuring the digital tools can be trusted.
The project has three main 'prongs':
First, the predictive capabilities of state-of-the-art computational flow-physics modeling needs to be established, i.e., where can we trust computational fluid dynamics (CFD) simulations and where can we not? As component and sub-system design and analysis are increasingly dependent on CFD, methods for quantifying the flow regimes (and boundary conditions) where CFD can be trusted needs to be rigorously ascertained. Techniques combining multi-fidelity modeling and machine learning underpin the research methods here.
Second, there has to be a greater effort to leverage existing sensor network data from engines and aircrafts. Rather than treating the sensors and their data in isolation, there needs to be a more cohesive and unified instrumentation-based model that can be used for prognosis, diagnosis, risk mitigation and performance assessment. To fill in the blanks (sensors networks in aeronautics are in general sparse), machine learning tools will be paired with data from the sensors and a physics-rooted understanding of the system.
Third, the development of new methodological thrusts within the remit of Gaussian process regression, deep neural networks, subspace-based dimension reduction, polynomial multi-fidelity approximation and classification will be required. These machine learning ideas are pivotal to achieving the aforementioned objectives - necessitating both the generation of new theoretical ideas whilst ensuring their practical computational implementation - aimed at more physically accurate digital twins in aeronautics.
Application areas include: aircraft engines, airplane wings and micro air vehicles (drone).