Abstract:In the recent decade, the reliability analysis of a stochastic system coupled with the uncertainty related to the system's parameter has attracted much attention. Probability density evolution method (PDEM) is one of the viable options that estimates the probability density function of the structural response by solving generalized density evolution equations (GDEEs). The advantage of PDEM is that it is derived based on the principle of probability conservation, where GDEEs are decoupled from the physical system. In general, GDEEs in PDEM are solved using a finite difference scheme in which the accuracy of the numerical solution depends on the number of temporal and spatial discretizations , leading to computationally inefficient for high-fidelity models. With this in view, this study proposes a physics-informed neural network (PINN), a novel deep learning method, based PDEM, for solving the GDEEs. PINN utilizes physical information in the form of differential equations to enhance the performance of the neural networks. This method does not need any interpolation or coordinate transformation, which is often seen in any numerical scheme, thus the computational budget is reduced. Three numerical examples are presented in this study to illustrate the proposed PINN-based PDEM, including a Van-der-Pol oscillator subjected to Gaussian white noise , a one-storey moment resisting frame coupled with a nonlinear energy sink with negative stiffness and sliding friction, and a high-rise timber building coupled with shape memory alloy-based outriggers. The first example is utilized to show the accuracy of the proposed method by comparing results with the Fokker–Planck–Kolmogorov equation and Monte Carlo simulation. The rest two examples are investigated for estimating time-dependent probability of failure. Numerical results show that the proposed PINN-based PDEM can estimate the probability of failure efficiently.

Data-driven method to learn the most probable transition pathway and stochastic differential equations

A data-driven approach for discovering the most probable transition pathway for a stochastic carbon cycle system

Computing Transition Pathways for the Study of Rare Events Using Deep Reinforcement Learning

Transfer learning of phase transitions in percolation and directed percolation

Data-driven approximation for extracting the transition dynamics of a genetic regulatory network with non-Gaussian Lévy noise

Learning phase transitions by siamese neural network

Detecting the Most Probable High Dimensional Transition Pathway Based on Optimal Control Theory

Data-driven approach for extracting the most probable exit trajectory of stochastic dynamical systems with non-Gaussian Lévy noise

Learning noise-induced transitions by multi-scaling reservoir computing

Deep Reinforcement Learning in Finite-Horizon to Explore the Most Probable Transition Pathway

Neural Jump Stochastic Differential Equations

Reliability assessment of stochastic dynamical systems using physics informed neural network based PDEM

Learning interpretable dynamics of stochastic complex systems from experimental data

Trans-Net: A transferable pretrained neural networks based on temporal domain decomposition for solving partial differential equations

Statistical-Physics-Informed Neural Networks (Stat-PINNs): A Machine Learning Strategy for Coarse-graining Dissipative Dynamics

Supervised and unsupervised learning of directed percolation

Stochastic physics-informed neural ordinary differential equations

Learning nonequilibrium statistical mechanics and dynamical phase transitions

Adapting Physics-Informed Neural Networks for Bifurcation Detection in Ecological Migration Models

Evolutionary probability density reconstruction of stochastic dynamic responses based on physics-aided deep learning

Generalizable Physics-Informed Learning for Stochastic Safety-Critical Systems