Constitutive parameter calibration for structural steel: Non-uniqueness and loss of accuracy

Abstract

Finite element (FE) simulations used to characterize extreme limit states in steel structures require the calibration of numerous parameters. Calibration of these models for large strains (greater than 0.3 or so) cannot be performed using stress–strain curves on standard tests, since the stress state in these tests becomes non-homogenous due to necking or buckling which occur at lower strains. As a result, calibration is often performed by matching load–displacement curves of calibration specimens to those obtained through complementary FE simulations. In these situations, multiple parameter sets produce strain fields that match the measured load–displacement response, resulting in non-unique parameter fits. A series of 2400 FE simulations with 4 specimen geometries, 300 material parameters sets, and 2 loading histories indicates that multiple trial parameter sets produce excellent load–displacement match with the true material response, implying that the method is highly susceptible to non-unique fitting. All simulations use the Armstrong and Frederick constitutive model with a von Mises yield surface. The impact of non-unique fitting is assessed through FE simulations based on parameter sets that show excellent load–displacement match with calibration specimens. It is determined that the non-uniqueness does not significantly affect the prediction of peak force. However, it severely impacts the accuracy in prediction of internal plastic strains, with errors as large as 50% with respect to the true material. This has serious implications for FE simulation used to characterize extreme, strain-based limit states such as fracture. Strategies for mitigation of this inaccuracy are presented, along with limitations of the study.