上海科技大学知识管理系统

We propose an invertible neural network (INN) based approach for uncertainty propagation and Bayesian estimation in unknown dynamical systems. The INN is trained to learn the short time flow map of the system, enabling the prediction of long time system behavior from arbitrary initial conditions. For uncertainty quantification, the method efficiently estimates solution statistics under uncertain initial conditions. By leveraging the change of variable rule, it computes the probability density function (PDF) of the solution when the initial condition is described by a PDF. Furthermore, the approach integrates with Bayesian filtering to estimate system states using observational data. Numerical experiments demonstrate the efficiency of the proposed method for both uncertainty propagation and Bayesian estimation.