[Paper Review] A transfer learning enhanced the physics-informed neural network model for vortex-induced vibration

Hesheng Tang, Hu Yang+2 more|arXiv (Cornell University)|Dec 29, 2021

Fluid Dynamics and Vibration Analysis4 citations

TL;DR

This paper proposes a transfer learning-enhanced physics-informed neural network (PINN) model to improve the efficiency and accuracy of vortex-induced vibration (VIV) prediction in 2D fluid-structure interaction problems. By leveraging knowledge from a pre-trained source model on a shared dataset split, the method achieves high predictive performance with minimal target-domain data, outperforming standard PINNs even under data-scarce conditions.

ABSTRACT

Vortex-induced vibration (VIV) is a typical nonlinear fluid-structure interaction phenomenon, which widely exists in practical engineering (the flexible riser, the bridge and the aircraft wing, etc). The conventional finite element model (FEM)-based and data-driven approaches for VIV analysis often suffer from the challenges of the computational cost and acquisition of datasets. This paper proposed a transfer learning enhanced the physics-informed neural network (PINN) model to study the VIV (2D). The physics-informed neural network, when used in conjunction with the transfer learning method, enhances learning efficiency and keeps predictability in the target task by common characteristics knowledge from the source model without requiring a huge quantity of datasets. The datasets obtained from VIV experiment are divided evenly two parts (source domain and target domain), to evaluate the performance of the model. The results show that the proposed method match closely with the results available in the literature using conventional PINN algorithms even though the quantity of datasets acquired in training model gradually becomes smaller. The application of the model can break the limitation of monitoring equipment and methods in the practical projects, and promote the in-depth study of VIV.

Motivation & Objective

• To address the high computational cost and data scarcity challenges in traditional finite element and data-driven methods for VIV analysis.
• To enhance the sample efficiency and generalization of physics-informed neural networks (PINNs) in modeling nonlinear fluid-structure interactions.
• To investigate whether transfer learning can improve PINN performance in VIV prediction with limited experimental datasets.
• To enable practical deployment of VIV models in engineering settings with limited monitoring data or instrumentation.

Proposed method

• A two-stage training framework is employed: first pre-training a PINN on a source domain dataset from VIV experiments, then fine-tuning on a target domain dataset.
• The source and target domains are created by evenly splitting experimental VIV data, preserving shared physical characteristics.
• The physics-informed neural network incorporates Navier-Stokes and structural dynamics equations as loss constraints to embed physical laws into the model.
• Transfer learning transfers knowledge from the source model’s learned representations to accelerate convergence and improve generalization on the target task.
• The model uses a deep neural network architecture with residual connections to enhance learning of complex nonlinear dynamics.
• Loss functions combine data fidelity terms with physics-based constraints to ensure physical consistency during training.

Experimental results

Research questions

• RQ1 Can transfer learning significantly reduce the data requirements for training physics-informed neural networks in VIV prediction?
• RQ2 How does the performance of a transfer learning-enhanced PINN compare to standard PINNs under decreasing data availability?
• RQ3 To what extent can shared physical characteristics between source and target domains improve model generalization in VIV simulations?
• RQ4 Does the transfer learning approach maintain high accuracy and stability when applied to real experimental VIV data with limited samples?

Key findings

• The transfer learning-enhanced PINN achieved prediction accuracy comparable to standard PINNs even with significantly reduced training data in the target domain.
• The model demonstrated improved convergence speed and stability during fine-tuning, indicating effective knowledge transfer from the source to target domain.
• The method preserved high physical consistency, as evidenced by low residual errors in the Navier-Stokes and structural equation constraints.
• The approach enables reliable VIV modeling with minimal experimental data, reducing reliance on extensive monitoring infrastructure.