Data driven model adaptation for identifying stochastic digital twins of bridges

• Applicants: Dr.-Ing. Jörg F. Unger; Dr. Martin Weiser
• Subject Area: Architecture, Building and Construction History, Construction Research, Sustainable Building Technology
• Term: since 2022
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 501811638

The goal of the proposal is the development of methods to support the creation of digital twins of bridges based on simulation models. In particular, this should include an estimate of the quality of prognosis of simulation results. This estimate is obtained via Bayes inference procedures using both laboratory as well as monitoring data. A particular challenge in the development of a simulation model for bridges is that in many cases the modelling assumptions from the design differ from the real implementation. For this reason, an adaptive process is necessary in which additional physical effects are added or changed (boundary conditions, geometry of the reinforcement layers, constitutive laws) until the digital and the real twin have identical properties within prescribed tolerances. However, the comparison between measured data and simulation results cannot be done manually, on the one hand due to the high dimension of the data space (large number and different types of sensors), on the other hand also due to the large amount of data resulting from monitoring. Furthermore, due to the ellipticity of the underlying differential equations, local changes at one point in the model lead to changes of the global behavior, i.e. the localization of model bias only from the data is often not possible. Therefore, the goal of the project is to develop a semi-automatic approach. Based on an initial, physically motivated model, the parameters of the model are calibrated using the data. In addition, a possible model improvement is identified from the data. Different approaches based on Gaussian processes, random fields for material parameters or a sparse dictionary learning approaches will be compared. Based on this data-driven model bias, the modeler shall develop a physically motivated model improvement (possibly with additional parameters) until the simulation model is sufficiently accurate. A large number of calculations of the forward model (FE model of the bridge) is required to calibrate the model parameters. In order to improve the performance of the methodology, the forward models are replaced by metamodels. In particular, Gaussian processes and Bayesian physics-informed neural networks are used for this purpose. The methodology will be validated within the project on different examples with varying degrees of complexity, starting with virtual data with simple beam models (including faults), existing bridge demonstrators on our test site and, finally, with real data from the demonstrator in the SPP. Parallel to the implementation of the methodology, a continuous development of the interfaces to the other projects in the SPP and an integration into a common platform is intended.

• DFG Programme: Priority Programmes
• Subproject of: SPP 2388: Hundred plus - Extending the Lifetime of Complex Engineering Structures through Intelligent Digitalization
• Co-Investigator: Professor Phaedon-Stelios Koutsourelakis, Ph.D.