Model updating for digital twins using Gaussian process inverse mapping models

In engineering dynamics, model updating is typically applied to minimize the mismatch between a physical system and its digital twin. This paper proposes to use inverse mapping models, based on Gaussian Processes (GPs). The latter are trained offline using simulated data, enabling fast online updating of physically interpretable parameter values in first-principles-based nonlinear dynamics models. The GPs infer parameter values based on time-domain features measured on the real system. Additionally, GPs enables uncertainty quantification of the inferred parameter values. A nonlinear multibody model is used to illustrate the capability of this method to update parameter values, with high computational efficiency, and extract corresponding uncertainty measures.