DEVELOPMENT AND APPLICATIONS OF MACHINE/DEEP LEARNING TECHNIQUES IN FLUID DYNAMICS

In this thesis, the machine/deep learning techniques are improved and applied in fluid flow problems. Firstly, the conventional multilayer perceptron (MLP) networks are enhanced by the Gaussian radial basis function (RBF) with trainable centers and widths. Activated by Gaussian RBFs, the proposed MLP-RBF network is more accurate and efficient in nonlinear regression problems. Secondly, in view of the network structure, RBF approximation is deployed in each hidden layer of the new network (RBF-MLP-II), which exhibits stronger predictive capability than other counterparts. Thirdly, the prevailing physics-informed neural networks (PINNs) are applied to predict the fluid flows in both global and local domains. The results show better performance in PINNs than pure MLPs, and the RBF-activated PINNs are compared with the tanh-activated ones. Last, the automatic differentiation-based PINNs are improved by the finite difference scheme, and the proposed FD-PINNs are validated using incompressible isothermal and thermal flows.

Author dc:creator

• JIANG QINGHUA

• Fluid dynamics
• Radial basis function
• Physics-informed neural networks
• Artificial neural networks
• Deep learning
• Machine learning