Flow-Based Bayesian Updating for Finite Element Models in Digital Twins

This paper introduces a flow-based Bayesian updating framework for finite element (FE) models in digital twin (DT) applications, addressing the challenges of real-time parameter updating and uncertainty quantification. By integrating conditional normalizing flows (CNFs) with Sequential Monte Carlo (SMC) sampling, the proposed method eliminates the need for explicit likelihood derivation, significantly reducing computational costs while maintaining robust Bayesian inference. A case study on a fork component demonstrated the framework's effectiveness, achieving mean absolute percentage errors (MAPEs) below 4% for density and elastic modulus predictions. Elastic modulus exhibited higher accuracy and more reliable posterior distributions due to its stronger influence on system responses. Key findings highlight the importance of parameter sensitivity and prior selection in prediction accuracy, with uniform priors showing reduced efficiency when true parameters deviate from the prior mean. The methodology provides an efficient and scalable solution for real-time FE model calibration, offering broad applicability in engineering domains requiring dynamic parameter updates. Future work will focus on extending the framework to handle time-varying parameters and improving prior selection strategies to enhance robustness.