(413g) A Physics-Informed Neural Network for Simulation of Turbulent Flow with Mass Transfer

Turbulent flow with mass transfer is common in chemical processes. The traditional approaches for numerical simulation of turbulent mass transfer processes have been mainly based on either turbulent Schmidt number model or two-equation model. However, these approaches, commonly implemented in CFD (Computational Fluid Dynamics) codes, usually suffer from lack of model parameters in solving the complex PDEs (Partial Differential Equations) of the model. In this study, an approach for fast estimating fields of velocity, temperature and concentration of turbulent flow with mass transfer is proposed by applying the PINN (physics-informed neural network) method. The PINN framework developed in this study is supervised by sparse observed data. In the training procedure, simplified PDEs with unknown turbulent viscosity and diffusion coefficient serve as the physical information provider. The results of rigorous CFD simulation of water- fluorescein sodium system, which has been verified against experiment, are adopted for data generation and results reference. The efficiency and accuracy of results of PINN, NN and turbulent Schmidt mechanism model in solving the same turbulent mass transfer problem are compared. A satisfactory agreement between the physical field distributions obtained by the proposed PINN and the ones by the CFD model indicates that the PINN trained by the proposed method is effective for simulating processes of turbulent mass transfer with incomplete physical information. The performances of PINN and NN in dealing with the noisy data are compared, and therefore, the ability of combination with experimental results for PINN is demonstrated. Besides, we also improve the PINN framework with multi-fidelity physical information embedded, and the effectiveness is demonstrated by deducing the turbulent viscosity and turbulent diffusion coefficient at different coordinates.