Machine-learning based prediction of crash response of tubular structures
Emmanouil Sakaridis, Nikolaos Karathanasopoulos, Dirk Mohr
Abstract
This paper proposes a machine learning based methodology for predicting the buckling response of tubular structures. An extensive dataset of force-time curves is generated using a calibrated finite element model within a parametric space where buckling response is highly non-linear. Based on a fully connected neural network template, the machine learning hyper-parameters are determined and the resulting model is evaluated on a separate test set, with regard to maximum and average load and energy absorption errors. This evaluation shows a non-random error distribution which can be correlated with the physical properties of the structural collapse. To validate this assumption, a similar error analysis is conducted between finite element simulations with varying geometric imperfections. Evaluation of imperfection sensitivity reveals a similar error distribution and comparison of individual curves shows that errors made by the neural network model have a physical interpretation. These results indicate that the proposed machine learning based approach is capable of predicting the crushing response with a level of accuracy comparable to the errors that would be caused by a minor change in geometric imperfection.