Fast solution of 3D transport processes using a physics-informed neural network with embedded transfer learning

Abstract

Ultrapermeable membranes (UPMs) have widely application prospects in water treatment and desalination. However, current membrane module limits the performance of UPMs due to the intensified concentration polarization and membrane fouling. Therefore, optimal design of membrane module is significantly critical to increase boundary mass transfer, but is requiring for sequential solution of many three-dimensional (3D) multi-physics models involving fluid flow and mass transfer, which is time-consuming and computational expensive. Herein, we propose a physics-informed neural network with embedded transfer learning (PINN-TL) that enables efficient and accurate solution for the flow and mass transfer characteristics at various geometric structures within the desired parameter range through five computational fluid dynamics simulations. The simulated results indicate that the average predicted accuracy for velocity magnitude, pressure, and concentration obtained from the comparison between PINN-TL and the finite element method are 98%, 99%, and 99%, respectively. Moreover, under GPU acceleration, the computational efficiency of PINN-TL is approximately 10 times higher than that of the finite element method. The developed intelligent solver of PINN-TL can be further applied in the optimal design of membrane module for UPM systems or other optimization problems involving transport process.