Generalization techniques of neural networks for fluid flow estimation

We demonstrate several techniques to encourage practical uses of neural networks for fluid flow estimation. In the present paper, three perspectives which are remaining challenges applications machine learning dynamics considered: 1. interpretability machine-learned results, 2. bulking out training data, and 3. generalizability networks. For interpretability, we first two methods observe internal procedure networks, i.e., visualization hidden layers application gradient-weighted class activation mapping (Grad-CAM), applied canonical estimation problems—(1) drag coefficient a cylinder wake (2) velocity from particle images. It is exemplified that both approaches can successfully tell us evidences great capability learning-based estimations. then utilize some bulk data super-resolution analysis temporal prediction NOAA sea surface temperature sufficient with limited amount be achieved problems. The model also discussed by accounting inter/extrapolation considering wakes behind parallel cylinders. find various patterns generated complex interaction between cylinders reconstructed well, even test configurations regarding distance factor. paper significant step toward laminar turbulent