物理
动量(技术分析)
高斯分布
粒子(生态学)
采样(信号处理)
弹道
职位(财务)
雷诺数
统计物理学
质点速度
机械
阻力
跟踪(教育)
经典力学
边界(拓扑)
流量(数学)
代表(政治)
计算流体力学
航程(航空)
结束语(心理学)
数学分析
曲率
边值问题
高斯过程
搭配(遥感)
人工神经网络
压缩性
重要性抽样
高斯函数
流体力学
纳维-斯托克斯方程组
歧管(流体力学)
不可压缩流
算法
稳态(化学)
颗粒流
流速
核(代数)
质点位移
作者
Na Ye,JINGKUAN HUANG,Yuxuan Jin,Lijie Yin,Kezhen Qian,Dezhen Chen
摘要
Unresolved particle–fluid systems require coupled prediction of fluid motion, particle dynamics, and interphase momentum exchange. We develop bidirectionally coupled physics-informed neural networks (Bi-PINNs) based on partitioned state exchange. F-PINN reconstructs velocity and pressure fields governed by the incompressible Navier–Stokes equations with particle feedback, whereas P-PINN predicts particle velocity and recovers trajectories through differentiable time integration. The subnetworks form a closed loop by exchanging the local background-velocity proxy, particle state, and interphase force. Multiscale temporal Fourier encoding and refined collocation sampling in particle near-field and near-wall regions improve the representation of transient and localized disturbances. Faxén parallel-wall drag correction, annular background-velocity sampling, and discretely normalized Gaussian momentum feedback define the force closure for confined low-Reynolds-number transport. Architecture and sampling analyses show that lower physics losses do not necessarily imply greater predictive accuracy and that balanced regional sampling improves multivariable reconstruction. Verification against computational fluid dynamics solutions based on the same governing equations and closure model shows close agreement for the bulk flow, particle-induced disturbance, particle Reynolds number, and trajectory over 0–1 s. Across six independently trained cases, with the inlet centerline-to-initial-particle velocity ratio α ranging from 0.85 to 1.15, the mean terminal-trajectory relative error is 1.34%, with a maximum of 2.79%. The case-wise L2 relative errors in streamwise velocity and pressure range from 3.52 × 10−4 to 4.64 × 10−2 and from 3.42 × 10−3 to 1.18 × 10−2, respectively. These results demonstrate that the Bi-PINNs framework accurately reconstructs coupled particle–fluid dynamics without labeled data during optimization.
科研通智能强力驱动
Strongly Powered by AbleSci AI