.png){kind=icon}
.png){kind=icon}
Return
GatePINN: Adaptive gating and multi-scale frequency modulation for accurate stress prediction in turbulence
.png){kind=icon}
PRE
AI
Z
Z
T
X
DOI：10.1016/j.jcp.2026.114784.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Accurate stress prediction in turbulent flow simulations remains a significant challenge due to abrupt spatial variations and multi-scale frequency patterns that existing Physics-Informed Neural Networks struggle to capture effectively, particularly in regions characterized by high gradient changes and complex flow dynamics. To address these limitations, this study presents GatePINN, a novel neural network architecture specifically designed to enhance stress prediction accuracy by resolving spatial inconsistencies and multi-scale frequency challenges inherent in turbulent flow modeling. The proposed methodology integrates three key innovations: an Adaptive Weighted Gating Mechanism that dynamically adjusts loss function weighting based on local prediction errors, Adaptive Multi-Scale Frequency Modulation layers that utilize Fast Fourier Transform to capture complex frequency patterns, and a dynamic physics-informed loss weighting strategy that intensifies physical constraint enforcement proportionally to local prediction errors, ensuring enhanced adherence to conservation laws in challenging flow regions. The architecture prioritizes high-error regions while simultaneously processing both high- and low-frequency components through learnable frequency domain parameters. Comprehensive evaluation on the benchmark periodic hills flow and turbulent cylinder flow at Re=3900 demonstrates superior performance, with GatePINN achieving R2 exceeding 0.98 on both cases and significantly outperforming baseline methods. GatePINN successfully addresses some limitations in turbulent flow modeling by combining adaptive error-driven training with multi-scale frequency processing, thereby offering a promising approach for physics-informed neural networks in computational fluid dynamics applications.
Keywords:
Adaptive multi-scale frequency modulation
Physics-informed neural networks
Computational fluid dynamics
Reynolds stress prediction
Adaptive weighted gating mechanism
Journal
IF:
3.8
Papers:
1.5W
Citations:
7.4W
Organization
Cited Papers