Reliable knowledge of damage evolution and failure of materials is an important prerequisite of a design process. In this context, strain localisation analyses have been shown to be well suited for predicting the initiation of ductile fracture once properly calibrated. In previous studies from the authors, strain localisation analyses have been successfully used as a pre-processor step to calibrate a ductile fracture model, or as a post-processor step to determine the location and time of failure initiation in various materials and specimens. While strain localisation analyses are rather efficient when carried out as pre- or post-processor steps, their use in-situ finite element analyses is still not possible. Indeed, under arbitrary non-proportional load histories, one has to track the evolution of a very large number of imperfection bands for every element and time step of the finite element analysis. This renders the simulation highly inefficient, especially when considering explicit finite element methods. To overcome this limitation, we study the feasibility of using a surrogate model for strain localisation analysis to enable its use in-situ explicit finite element simulation. To build our surrogate model, we make use of an artificial neural network trained on data generated by strain localisation analyses. One of the outcomes of this study is that artificial neural networks trained using data from proportional loading can be used to predict the response for non-proportional loading as well.