Scholarlink

<p>Emerging as surrogate models in response simulations, the physics-informed neural networks (PINNs) are expected to replace computationally expensive time-history analysis methods and achieve rapid prediction of dynamic responses of nonlinear systems. For existing PINN models, the global equation of motion of the nonlinear system is directly incorporated into the loss function, requiring the networks to learn the global dynamic behavior of the system. As a result, network training becomes challenging and hinders the application of PINNs to complex nonlinear systems. In the present study, an impulse response function-based PINN, termed the IRF-PINN, is proposed to facilitate the practical application of PINNs in predicting the dynamic responses of engineering structures installed with nonlinear energy-dissipating devices. Instead of using the original nonlinear equation of motion, the impulse response functions (IRFs), which characterize higher-order physics information of the structure, are integrated into the loss function. As such, the linear dynamic behavior of the structure, incorporating the effects of energy-dissipating devices modeled by nonlinear restoring forces, can be automatically revealed. Consequently, only a lightweight network consisting of a temporal convolutional network (TCN) module and a long short-term memory (LSTM) module is required to predict the unknown restoring forces of the devices under external loadings. The remaining structural responses can then be readily derived from the predicted restoring forces. A 6-story shear-type structure with nonlinear viscous dampers is investigated by different PINN models, demonstrating the significantly improved training efficiency and prediction accuracy of IRF-PINN for nonlinear time-history response prediction compared with the conventional PINN model. Subsequently, a complex cable-stayed bridge installed with hysteretic isolators is explored using the proposed IRF-PINN model, and the Monte Carlo simulation (MCS) is conducted to evaluate the failure probability of the bridge subjected to seismic excitations, indicating the promising prospect of the present approach in engineering applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Impulse response function-based PINN for dynamic response prediction of engineering structures with nonlinear energy-dissipating devices

  • 期刊论文

  • 发表日期：

    2026年1月22日

查看全文
  • Jingwei Liang,
  • Cheng Su

摘要

Emerging as surrogate models in response simulations, the physics-informed neural networks (PINNs) are expected to replace computationally expensive time-history analysis methods and achieve rapid prediction of dynamic responses of nonlinear systems. For existing PINN models, the global equation of motion of the nonlinear system is directly incorporated into the loss function, requiring the networks to learn the global dynamic behavior of the system. As a result, network training becomes challenging and hinders the application of PINNs to complex nonlinear systems. In the present study, an impulse response function-based PINN, termed the IRF-PINN, is proposed to facilitate the practical application of PINNs in predicting the dynamic responses of engineering structures installed with nonlinear energy-dissipating devices. Instead of using the original nonlinear equation of motion, the impulse response functions (IRFs), which characterize higher-order physics information of the structure, are integrated into the loss function. As such, the linear dynamic behavior of the structure, incorporating the effects of energy-dissipating devices modeled by nonlinear restoring forces, can be automatically revealed. Consequently, only a lightweight network consisting of a temporal convolutional network (TCN) module and a long short-term memory (LSTM) module is required to predict the unknown restoring forces of the devices under external loadings. The remaining structural responses can then be readily derived from the predicted restoring forces. A 6-story shear-type structure with nonlinear viscous dampers is investigated by different PINN models, demonstrating the significantly improved training efficiency and prediction accuracy of IRF-PINN for nonlinear time-history response prediction compared with the conventional PINN model. Subsequently, a complex cable-stayed bridge installed with hysteretic isolators is explored using the proposed IRF-PINN model, and the Monte Carlo simulation (MCS) is conducted to evaluate the failure probability of the bridge subjected to seismic excitations, indicating the promising prospect of the present approach in engineering applications.