A new framework incorporating long short-term memory and physics-informed neural networks for solving transient fluid flow problems in turbine vanes

Accurately and efficiently solving transient fluid flow problems in turbine vanes is of great importance for the design as well as the optimization of turbines. Numerical simulation has played an important role in solving transient fluid flow problems, which is accurate but time-consuming for repeated calculation in an optimization process. To address the issue in computational efficiency, this paper proposes a new framework which incorporates the long short-term memory (LSTM) with physics-informed neural network (PINN), abbreviated as LS-PINN, to solve transient fluid flow problems. Compared with purely data-driven methods, the LS-PINN framework integrates physical information into training process, which can not only improve the prediction accuracy but also enhance the interpretability of the results. Meanwhile, the LS-PINN framework enables quick predictions of multiple physical quantity fields in turbine vanes, with varying gas inlet angles and limited data, by employing transfer learning. Examples and results demonstrate feasibility and high accuracy. The transfer learning-enhanced LS-PINN framework achieves comparable accuracy to the original framework, with only one-third of the training data and the computational time, demonstrating significant potentials in engineering applications. In addition, the generalization ability of the LS-PINN framework on time and related parameters are studied. • A new framework is established for solving fluid flow problems in turbine vanes. • The framework innovatively introduces a two-stage training method. • The proposed framework demonstrates feasibility and high accuracy. • The proposed framework could accurately reconstruct the entire fluid flow processe. • Transfer learning is utilized and higher efficiency can be obtained. • The framework has good generalization ability for time and parameters.