Machine Learning based Digital Twin Simulation and Prediction of Structural Component's Fatigue Life

Abstract

A key component for implementing the digital twin approach is to apply a validated high fidelity simulation tool to generate a mapping between the virtual test and structural performance. Due to the high computational intensity of physical simulation tools, their application for a complex system along with its error estimation can be time consuming. In addition, given the limited data gathered from sensors, onsite inspection, and tests at different configurations, it is imperative to create a high fidelity and efficient model based on the previously gathered information and enhance the model when more data points are gathered. Motivated by this, we develop a machine learning based digital twin simulation framework to predict fatigue life of a structural component from available information gathered. Different from a conventional physical simulation approach, the prediction error from the physical simulation and machine learning are explicitly obtained, in addition to the improvement of the computational efficiency at the prediction stage via the trained machine learning model. To illustrate the idea of this modeling strategy, we applied our developed 3D extended finite element toolkit for Abaqus (XFA3D) as a virtual testing tool for fatigue crack path and life prediction of a welded metallic component in conjunction with the observed testing data. Using machine learning techniques, we first estimate prediction error from the physical model based on previous cross validation results, and then predict the fatigue life in the presence of uncertainties associated with fabrication induced imperfection, welding induced residual stress, and the machine learning errors. It is found that the inclusion of as-manufactured characteristics and uncertainties are essential for the application of a digital twin approach for the total life management of aging structures.