Predicting Buckling and Plasticity of Finite Element Models Using Machine Learning

Abstract

The ongoing demand for bigger and more efficient ships pushes their designs towards the strength limits. Sometimes, ship structures are pushed beyond their limits with the possibility of significant negative economic and environmental impact or, in the worst case, impact on human life. This makes it explicitly clear why the development of accurate methods and models is still required to predict if a design can withstand the expected loads during its operations. In the past two decades, Machine Learning (ML) has been applied in the field of structural design. Still, the amount of research regarding the scalability and generalizability of ML within structural design is limited. This research aims to investigate the use, scalability, and generalizability of ML to predict the ultimate strength of stiffened panels. A large dataset of stiffened panels loaded under longitudinal uni-axial compression and lateral pressure is generated. This set is based on a representative geometrical parameter set in shipbuilding obtained by the optimization of an analytical buckling model for stiffened panels. This model incorporates several nonlinearities such as initial deflection, large deflection theory, and residual stress. Finite Element Analysis (FEA) is used to obtain the ultimate strength and stress distribution at the moment of failure for every stiffened panel. This data is used to train two Convolutional Neural Networks (CNN) to predict the ultimate strength and stress distribution at the moment of failure. Overall, the CNNs are an excellent tool for these predictions when sufficient data is available, the data shows substantial geometrical similarity to the training data, and when the data is within the scope of training. When predicting the ultimate strength of the stiffened panels in this research, it is required to have at least 4000 data samples to obtain a well-trained network. The scalability of the ultimate strength predictions produces acceptable results within the 5% margin of the training scope. Unstable prediction behavior is observed when predicting stress distributions. Both the ultimate strength and stress predicting models are also trained on stiffened panels curved in the transverse direction to test the generalizability of the CNNs. Poor predictive capabilities were observed when predicting the ultimate strength and stress distributions of the curved plates. The use of ML in structural engineering proves a useful concept and provides multiple opportunities for further research.