Improvement of beam type element models of welded tubular junctions for fatigue analysis using artificial neural networks

The use of numerical methods for structural analysis has been increasingly integrated within the design process in different engineering fields over the last decades, inasmuch as the capacity of the computing resources have growth. This gave rise to calculation techniques based on virtual models such as the finite element method, which is nowadays a reference method for evaluation of complex tubular structures with vast application in the industry. For such type of structures, modeling approaches based on beam type elements are usually employed since they provide simplicity and low computational costs. Nevertheless, these elements have the drawback that they cannot account for local geometric characteristics and therefore consider and strain concentrations consequence of the local joint geometry. These local strains are of special concern in welded junctions subjected to fatigue loads, since are the ones that will most probably lead to failure. Consequently, improving beam type elements takes special relevance. In this scenario, the present paper evaluates a novel methodology to improve strain results of beam element type models of welded T-junctions using artificial neural networks to predict the correction values depending on the junction geometry and load type. Detailed validated models are used as reference for network training. The paper first evidence the importance of the adequate selection of the training data set in the network precision and a methodology to ensure the best network selection is described. Then, the network capability to correct beam element type deviations is demonstrated. The obtained results show that the aid of neural networks to finite element beam T-junctions models can improve local strain result deviations by more than 90 % in most cases, which potentially allows performing fatigue analysis using this simplified modelling technique.