Inverse surrogate model for deterministic structural model updating based on random forest regression

This article presents a novel method for deterministic finite element model updating that is based on an “inverse surrogate model”. The latter is a regression model that uses structural responses as independent variables and structural properties as dependent ones. Such regressor is trained on structural responses using a finite element model having in input the structural properties to be updated. The training set is obtained by executing multiple finite element model instances where each instance is formed by assigning random, though realistic, values to the structural properties. Experimentally measured structural responses can then be given as input to the trained “inverse surrogate model”, and the structural properties, which are used to update the finite element model, are obtained as output. Random forest algorithm is identified as a good candidate to perform this type of regression, and the performance offered by this method is compared to that of the “standard” finite element model updating based on the particle swarm optimization algorithm on both a numerical and an experimental case study where the structural responses are represented by a set of modal parameters. The results show that the proposed approach offers higher accuracy on the estimation of the structural properties, at the cost of a slightly inferior matching of the modal parameters. This indicates that the “inverse surrogate model” is less susceptible to the ill-conditioning issue that plagues optimization-based finite element model updating, where similar structural responses can arise from various combinations of structural properties. The results also show a good repeatability and that the computational costs are similar or better than those of the particle swarm optimization algorithm.