Experimental application of neural operators for prediction of bluff body wakes

Capable of learning to map between infinite-dimensional functional spaces, neural operators are an exciting and powerful machine learning toolset. The underlying principles of these methods, first applied by Lu et al. (2019) as the Deep Operator Network (DeepONet), have been shown to enable data-driven time-efficient solvers for families of partial differential equations. Fourier Neural Operators (FNOs), introduced by Li et al. (2020), have been previously shown able to learn accurate solution operators based on synthetic data generated by the Navier-Stokes equation. Once trained, FNOs can produce full-field approximations in just milliseconds. In this talk we apply FNOs to experimental flow data, trying to predict the temporal development of several bluff-body wake configurations. We also explore the generalizability of the learned operator networks by testing the performance of fully trained FNOs on wake configurations different from those used for training.