Coupling machine learning and synthetic image DIC-based techniques for the calibration of elastoplastic constitutive models

Abstract

Today, most design tasks are based on simulation tools. However, the success of the simulation depends on the accurate calibration of constitutive models. Inverse-based calibration methods, such as the Finite Element Model Updating and the Virtual Fields Method, have been developed for identifying constitutive parameters. These methods are based on mechanical tests that allow heterogeneous strain fields under the “Material Testing 2.0” paradigm in which digital image correlation plays a vital role. Although these methods have been proven effective, constitutive model calibration is still a complex task. A machine learning approach is developed and implemented to calibrate elastoplastic constitutive models for metal sheets, using datasets populated with finite element simulation results of strain field data from mechanical tests. Feature importance analysis is conducted to understand the importance of the different input features and to reduce the computational cost related with model training. Synthetic image DIC-based techniques were coupled with the numerically generated database, enabling the construction of a virtual experiments database that accounts for sources of uncertainty that can influence experimental DIC measurements. A robustness analysis of the methodology is performed for the boundary conditions of the test.