Machine Learning of Structural Aerodynamics: Physics-enhanced Data-Driven Modelling

• Applicant: Dr.-Ing. Igor Kavrakov
• Subject Area: Applied Mechanics, Statics and Dynamics ; Structural Engineering, Building Informatics and Construction Operation ; Fluid Mechanics
• Term: since 2021
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 491258960

An abundant amount of data gathered during wind tunnel testing and health monitoring of structures inspires the use of machine learning methods to replicate the aerodynamic forces acting on lifeline structures such as long-span bridges, tall towers, and masts. These forces are critical for structural design and sustainability, and their modelling can be challenging due to their nonlinear and nonstationary nature. Traditionally, semi-analytical and Computational Fluid Dynamics (CFD) models have been utilised as physics-based models to replicate the aerodynamic forces. Data-driven aerodynamic models have recently started to gain attention in wind engineering due to their flexible mathematical formulation compared to the semi-analytical models, computational efficiency compared to the CFD models, and the potential for continuous model updating. However, data-driven models are black-box models that may be prone to robustness issues due to a lack of physical insight.The project proposes a physics-enhanced data-driven methodology for modelling the aerodynamic forces acting on civil structures. The methodology shall employ Gaussian Processes (GPs) as a machine-learning method to model the nonlinear aerodynamic forces, including a semi-analytical model as prior physics-based knowledge. Several mathematical formulations of the data-driven model shall be explored to determine their adequacy of capturing nonlinear features in the aerodynamic forces, such as higher-order harmonics and nonstationarity. In such a formulation, the output is the aerodynamic force, while the input is the free-stream turbulence, structural motion, and potentially autoregressive force terms. Leveraging a semi-analytical model as a prior distribution of the GPs incorporates physics into the model, thereby increasing the model robustness. A strategy for generating an input training signal shall be developed to train the data-driven model adequately. Based on this strategy, CFD data shall be generated to train the data-driven model and then verify its capabilities to predict multiple aerodynamic phenomena such as buffeting, vortex-induced vibration, and flutter, including post-flutter behaviour. The proposed data-driven model shall be verified against benchmark analytical models and validated against wind tunnel experiments. As a final part of the methodology, an online learning algorithm is proposed for updating the data-driven model based on a continuous stream of data that is typically collected during structural health monitoring.Finally, the proposed methodology aims to synergise knowledge from data and physics to provide a robust and efficient nonlinear aerodynamic model. Such a model would offer a deeper understanding of the physical processes during fluid-structure interaction and an accurate prediction of the aerodynamic forces for the design and monitoring of structures. Thus, it has the potential to be of relevance for academics as well as practitioners.

• DFG Programme: WBP Fellowship
• International Connection: United Kingdom
• Host: Professor Dr. Allan McRobie