Shallow Neural Network Predictions of Unsteady Flows About a Rotationally Oscillating Cylinder

摘要

This study aims to accurately predict the unsteady time response of a rotationally oscillating cylinder in laminar flow using a shallow neural network. Of particular interest is determining the regions where the vortex shedding frequency is identical to the excitation frequency of the cylinder (lock-on) and the region where multiple frequencies exist (non lock-on). Determining these regions is computationally expensive as high fidelity CFD simulations need to be completed for a large number of excitation conditions. This study shows that the computational cost can be dramatically reduced by using shallow neural networks to "predict" the unsteady time response as well as lock-onon lock-on regions, while retaining a high level of accuracy. For training, the unsteady time response signal is first transformed into the frequency domain, where a small number of dominant Fourier modes are identified. The network was then trained in the frequency domain to reduce the computational cost and to alleviate issues with over-fitting. The predicted frequency domain solutions were then transformed back into the time domain for analysis. The results indicate that the current approach generates moderate fidelity predictions at a fraction of the computational time required by high-fidelity simulations.