物理信息神经网络在两相流中的应用

Application of physics-informed neural network in two-phase flow

机器学习和数据科学相关研究从计算机科学学科涌向化学工程,将为化学工程领域创造变革范式的机会,其中物理信息神经网络(PINN)因将物理方程嵌入神经网络中使得网络输出满足物理规律而获得广泛关注。首先介绍PINN的算法思想及其采样策略;进一步讨论对PINN的损失函数不同的处理方式,主要包括无观测值、方程降阶、方程离散化和只嵌入部分物理方程等;最后概述了PINN方法在气液两相流、多孔介质两相流、液固两相流、两相流传热等领域最新进展。

The influx of research on machine learning and data science from the field of computer science into chemical engineering presents transformative opportunities for chemical engineering paradigms.Among them,physics-informed neural network(PINN)has gained wide attention because it embeds physical equations into neural networks so that the network output satisfies physical laws.This work begins by introducing the algorithm ideas and sampling strategies of PINN.It further discuss various treatment of the PINN loss function,mainly including cases with no observational data,equation reduction,equation discretization,and partial embedding of physical equations.Finally,it provides an overview of recent progress in the application of PINN to areas such as gas-liquid two-phase flow,two-phase flow in porous media,liquid-solid two-phase flow,and heat transfer in twophase flow.