Millionfold accelerated AI solver for 3D multi-physical simulations of ultrapermeable membranes

Abstract

Solving three-dimensional (3D) multi-physics forward and inverse problems is indispensable for fundamental understanding and optimal design of membrane-based desalination systems. Unfortunately, it is computationally expensive when applying traditional numerical methods. Herein, a modified Fourier neural operator (FNO)-based method is proposed to efficiently solve complex 3D multi-physics problems. The intelligent solver solves the 3D forward problems in seconds, which is approximately 10⁵-10⁶ times faster than traditional finite-element based method with a comparable solution quality. The average prediction accuracy is more than 96%. Moreover, the proposed FNO-based method is mesh-independent and has zero-shot super-resolution ability. It can be used to provide a fast solution for the optimal design of membrane module to mitigate concentration polarization and membrane fouling for next-generation ultrapermeable membrane system.