Crack assessment for pipe components of a nuclear power plant or oil/gas pipeline is one of the essential procedures to ensure safe operation services. To assess cracked pipes, J-integral has been considered as a theoretically robust and useful elastic-plastic fracture parameter, so that the estimations of J-integral for various pipe geometries, material properties and loading conditions are highly needed. For this reason, many engineering predictive solutions for J-estimations based on finite element (FE) analyses have been developed. Generally, many engineering predictive solutions have been suggested as a tabular-form or closed-form. Among them, the closed-form solution is more preferred than a tabular-form solution for its convenience when many lots of interpolation are required to use it. However, the accuracy of the closed-form solution tends to be significantly reduced as the number of design parameters increases. Moreover, since there is no strict rule to define the form of functions as well, the accuracy of the closed-form solution is inevitably dependent on the rule of thumb. Therefore, it is highly required to suggest a new approach for J-estimation of cracked pipes with various geometries, material properties and loading conditions.
In this paper, we propose an efficient approach based on a machine learning technique to estimate J-integral for surface cracked pipes with various geometric sizes and material properties under axial displacement loading condition. Firstly, parametric FE analysis studies were systematically performed to produce the coefficients representing the engineering J-estimation for the corresponding cracked pipe. Secondly, artificial neural network (ANN) models based on deep multilayer perceptron technique were trained based on FE results. The five input neurons (pipe geometries and material properties) and the two output neurons (the coefficients representing the engineering J-estimation) were considered. Lastly, the accuracy of the trained ANN model was studied by comparing to that of the closed-form solution from multi-variable regressions.