Simulation models are an essential tool to support decision making in civil engineering. Since the structures are often unique, full scale experimental validations are very costly or even impossible and the consequences of failure are severe, the simulation models have to be reliable and should provide results with sufficient safety margins, often resulting in very conservative designs. For innovative approaches such as 3D concrete printing, these models are even more important - supporting not only an a priori design optimization, but provide the option for a real-time process control or a quality assessment in a post-processing by evaluating the impact of discrepancies between the as-designed and as-built structures on quantities of interest. The simulation models are typically parameterized, and with the inclusion of noisy measurement data and prior knowledge of model parameters, Bayesian methods allow for the estimation of posterior parameter distributions. However, engineers are often interested in quantities of interest (QoIs) that are either not measured at the location of interest or cannot be measured directly, such as maximum stresses, failure probabilities or remaining useful life. Propagating posterior parameter distributions to these QoIs generally underestimates their uncertainty, especially when dealing with large datasets which is typical for digital twins with a constant data flow. One of the reasons is the fact that model form uncertainty is usually neglected. However, for complex structures involving interacting physical processes, model form uncertainty due to modeling assumptions and simplifications is unavoidable to ensure computational tractability of the simulation models. As a consequence, it is required to incorporate these additional uncertainties in the computational approach and model updating procedure. This work presents embedded Bayesian methods that integrate model form uncertainty into the parameter space by transforming the deterministic model into a stochastic one [1]. The propagation of these parameter distributions through the model is achieved using polynomial chaos expansion techniques. Additionally, the likelihood is redefined using Approximate Bayesian Computation (ABC) methods. The proposed methods are applied to several structural engineering case studies, ranging from model calibration for 3D concrete printing based on laboratory tests to the digital twin of the Nibelungen bridge in Worms, Germany.