Aeroelastic force prediction via temporal fusion transformers

Abstract

Aero-structural shape design and optimization of bridge decks rely on accurately estimating their self-excited aeroelastic forces within the design domain. The inherent nonlinear features of bluff body aerodynamics and the high cost of wind tunnel tests and computational fluid dynamics (CFD) simulations make their emulation as a function of deck shape and reduced velocity challenging. State-of-the-art methods address deck shape tailoring by interpolating discrete values of integrated flutter derivatives (FDs) in the frequency domain. Nevertheless, more sophisticated strategies can improve surrogate accuracy and potentially reduce the required number of samples. We propose a time domain emulation strategy harnessing temporal fusion transformers (TFTs) to predict the self-excited forces time series before their integration into FDs. Emulating aeroelastic forces in the time domain permits the inclusion of time-series amplitudes, frequencies, phases, and other properties in the training process, enabling a more solid learning strategy that is independent of the self-excited forces modeling order and the inherent loss of information during the identification of FDs. TFTs’ long- and short-term context awareness, combined with their interpretability and enhanced ability to deal with static and time-dependent covariates, make them an ideal choice for predicting unseen aeroelastic forces time series. The proposed TFT-based metamodel offers a powerful technique for drastically improving the accuracy and versatility of wind-resistant design optimization frameworks.

Omar A. Mures

Instructor

My research interests include Deep Learning, Computer Vision and Computer Graphics.