Reconstruction of ship propeller wake field based on self-adaptive loss balanced physics-informed neural networks.

This paper explores the application of Physical-informed Neural Networks (PINNs) for reconstructing ship propeller wake fields, introducing a novel approach known as self-adaptive loss balanced physics-informed neural networks (LB-PINNs). It can enhance the accuracy of the reconstruction process by incorporating an adaptive weight to balance the loss term within the network. The initial sections of the paper present the foundational principles and framework of PINNs. To validate the efficacy of PINNs in solving partial differential equations, the well-known Burgers equation is applied and then these results with those obtained through LB-PINNs are compared. This comparative analysis highlights the superior performance of LB-PINNs in achieving accurate predictions. Moving forward, the open water characteristics of the KVLCC2 propeller are simulated by using computational fluid dynamics (CFD) software STAR CCM+, and the flow field information of the KVLCC2 propeller in open water is obtained. This simulation provides crucial flow field information, forming the basis for constructing a training sample set to train the neural network. The trained PINN and LB-PINN are used to infer approximate solutions of the governing equations at arbitrary time and space coordinates. The velocity and pressure distributions obtained by PINN and LB-PINN were compared with those simulated by STAR CCM+. The results confirm the applicability of both PINN and LB-PINN in the reconstruction of propeller wake fields, with LB-PINN demonstrating superior performance. • A novel method for solving governing differential equations of fluid mechanics based on LB-PINN is introduced. • LB-PINN's adaptive weight balances the loss term, reducing mean square error by an order of magnitude compared to PINN. • Sparse numerical simulation data suffices for LB-PINN training, enabling ship propeller wake reconstruction. • Comparing flow field reconstructions using PINN, LB-PINN, and CFD shows LB-PINN's validity and accuracy. [ABSTRACT FROM AUTHOR]