Predicting Non-Stationary Homogeneous Variable-Density Turbulence Using Taylor-Net

Deep learning has shown the potential to significantly accelerate the numerical simulation of fluids without sacrificing accuracy, but prior works are mostly limited to stationary flows with uniform density. In real-world engineering applications, turbulent flows are mostly three-dimensional, non-stationary, and have variable-density. Here we propose Taylor-Net, a hybrid model that combines deep neural networks with the numerical Taylor series method for 3D turbulent flow prediction. Across flows with different density-ratio, our method is over 3 orders of magnitude faster than high-fidelity numerical simulations. It also achieves higher accuracy than several strong physics-informed deep learning baselines. Most importantly, the predictions of our Taylor-Net pertain consistent physical characteristics including mass conservation and turbulent energy spectrum.