Damage Detection using Physics-based modeling and data-driven optimization

Environmental conditions, variability in building material, and fabrication error are several contributors that shorten a structure's service life. While infrastructure might not have visible exterior damage, there is always the risk for failure due to internal or obscured damage. There has been growing research in the area of Nondestructive testing. However, these methods often disregard uncertainties in the construction and maintenance processes. Nevertheless, providing an accurate and robust method for detecting, localizing, and estimating the severity of existing damage relies on parameterizing uncertainties surrounding material properties, composition, and loading history. This work combines conventional physics-based modeling with data-driven search strategies to infer unobservable information about the structure. The proposed approach uses Finite Element Analysis with automated adjoint calculation to efficiently obtain gradients in the solution with respect to the model parameters. The hybrid model updates the postulated material properties to match the mechanical compliance of a simulated structure with observations from a reference structure. Here we use a simple cantilever beam model to study and evaluate the effectiveness of the proposed approach. In addition, we analyze the results from multiple search strategies used to obtain the mechanical properties. We find that the benefit in search efficiency gained from the analytical evaluation of the gradients is offset by the additional computational cost of adjoint simulation compared to gradient-free search. However, we expect that the benefit of adjoint simulation would be more pronounced in problems with more degrees of freedom.