Dynamic model updating through reliability-based sequential history matching

Computer models enable the study of complex systems and are extensively used in fields such as physics, engineering, and biology. History Matching (HM) is a statistical calibration method that accounts for various sources of uncertainty to update model parameters and align output with observed data. By iteratively excluding regions of the parameter space unlikely to yield plausible outputs, HM identifies and samples from the so-called non-implausible domain. However, a limitation of HM is that it does not yield full Bayesian posterior distributions for model parameters. Moreover, HM requires re-execution from scratch when new data is observed, lacking the ability to leverage prior results. To address these limitations, we propose integrating sequential Monte Carlo (SMC) methods with HM to achieve full Bayesian posterior distributions for sequential calibration. The SMC framework offers a flexible and computationally efficient means to update previously constructed distributions as new data becomes available. This approach is demonstrated using an engineering example and a cardio-respiratory case study with sequential data. Our results show that small perturbations to the posterior distributions can be effectively learned sequentially by updating computed posterior distributions through the SMC framework, thereby enabling dynamic and efficient model updating for evolving data streams. • Efficient updating of data using history matching with sequential Monte Carlo. • Full posterior distributions for model parameters are estimated. • Reliability-based discovery and exploration of the non-implausible space is achieved. • Promising results for dynamic model updating.