A data-driven model of the yield and strain hardening response of commercially pure titanium in uniaxial stress

This study presents a technique to develop data-driven constitutive models for the elastic-plastic response of materials, and applies this technique to the case of commercially pure titanium. The complex yield and strain hardening characteristics of this solid are captured for random non-monotonic uniaxial loading, without relying on specific theoretical descriptions. The surrogate model is obtained by supervised machine learning, relying on feed-forward neural networks trained with data obtained from random loading of titanium specimens in uniaxial stress. Uniaxial tests are conducted in strain control, applying random histories of axial strain in the range [−0.04, 0.04], to prevent the occurrence of significant damage. The corresponding stress versus strain histories are subdivided into a finite number of increments, and machine learning is applied to predict the change in stress in each increment. A suitable architecture of the data-driven model, key to obtaining accurate predictions, is presented. The predictions of the surrogate model are validated by comparing to experiments not used in the training process, and compared to those of an established theoretical model. An excellent agreement is obtained between the measurements and the predictions of the data-driven surrogate model.