A novel deep-learning-based objective function for inverse identification of material properties

Abstract

Inverse methods are critical in exploring the constitutive models of materials. In the traditional finite element model updating method, an objective function is often established in the form of mean square error, which is simple and easy to use. However, this approach does not fully consider the prior knowledge of the material, leading to low computational stability and sensitivity to noise. To address these limitations, a novel objective function that considers deep semantic information and mechanical constraints was proposed based on deep learning techniques in this study. Compared with the traditional method, the proposed objective function can significantly reduce the effect of noise on inversion performance without any loss in computational efficiency. The method was validated through simulated experiments and successfully applied to the inverse identification of damage properties of real nuclear graphite materials under simple and complex stress states. The results demonstrated that the proposed objective function can significantly enhance the stability of the finite element model updating method, leading to a rapid convergence with high accuracy.